- Compact style
- Indico style
- Indico style - inline minutes
- Indico style - numbered
- Indico style - numbered + minutes
- Indico Weeks View
Please refer here for how to cite contributions. We sincerely thank all the authors and reviewers for their work on this volume: your contributions are much appreciated.
ACAT 2022 is in Bari, Italy in October 2022. The spring 2024 ACAT workshop is in preparation. We hope to see you again at the next ACAT workshop.
The 20th edition of ACAT will bring together experts to explore and confront the boundaries of computing, automated data analysis, and theoretical calculation technologies, in particle and nuclear physics, astronomy and astrophysics, cosmology, accelerator science and beyond. ACAT provides a unique forum where these disciplines overlap with computer science, allowing for the exchange of ideas and the discussion of cutting-edge computing, data analysis and theoretical calculation technologies in fundamental physics research.
ACAT promises to showcase an excellent set of plenary speakers, including (as highlight) Joseph Lykken (Fermilab, on Quantum Computing), Lenka Zdeborova (EPFL, on a Theory of Deep Learning), Barry C. Sanders (University of Calgary, on Quantum Machine Learning), Michael Spannowsky (Durham University, on Unsupervised Machine Learning for New Physics Searches), and Julia Fitzner (WHO, on WHO's Data Analysis Challenges during COVID-19 Pandemic).
There is a fundamental shift occurring in how computing is used in research in general and data analysis in particular. The abundance of inexpensive, powerful, easy to use computing power in the form of CPUs, GPUs, FPGAs, etc., has changed the role of computing in physics research over the last decade. The rise of new techniques, like deep learning, means the changes promise to keep coming. Even more revolutionary approaches, such as Quantum Computing, are now closer to becoming a reality.
To complement the plenary program, we invite you to submit abstracts for parallel talks and posters. The poster sessions will be provided through a virtual platform, allowing for a more convenient and exciting conference experience. Please join us to explore these future changes, and learn about new algorithms and ideas and trends in scientific computing physics. Most of all, join us for the discussions and the sharing of expertise in the field.
ACAT2021 will be about the frontiers and limits of AI and explore how we can develop sustainable techniques that can be widely used on a diverse range of computing architectures, in particular to exploit co-processors and High Performance Computing (HPC) facilities, to optimise the processing and analysis of scientific data.
As often has happened in the development of AI, theoretical advances in the domain have opened the door to exceptional perspectives of application in the most diverse fields of science, business and society at large. Today the introduction of Machine (Deep) Learning is no exception, and beyond the hype we can already see that this new family of techniques and approaches may usher a real revolution in various fields.
However, as it has happened time and again in the past, we start realizing that one of the limitations of Deep Learning is sheer computing power. While these techniques allow to tackle extremely complex problems on very large amount of data, the computational cost, particularly of the training phase, is rising fast. The introduction of meta-optimizations such as hyper-parameter scans may further enlarge the possibilities of Machine Learning, but this in turn will require substantial improvements in code performance.
At the same time, HPC is also in full evolution, offering new solutions and perspectives, both in hardware and in software. In this version of ACAT we would like to focus on how this renewed momentum in the HPC world may provide the necessary power to fulfill the revolutionary promises offered by recent breakthroughs in AI at large and in Machine Learning in particular.
Fierce competitions between major players such as Google and IBM signals the advent of quantum computing. Research is being actively pushed to apply quantum computing tools to solve the problem which would have been too difficult to obtain the outcome using conventional methods. Are we seeing quantum supremacy? If yes, then in how many layers of sophistication?
You can attend the workshop either via the virtual environment or via the local meeting (pending the government regulation). The local gethering will take place at IBS Science Culter Center (English language button at top right) in Daejeon, Republic of Korea (South Korea).
Daejeon is a business city located at the center of South Korea, and it is a seat of a South Korean government complex. Its history goes back to thousands of years ago, when prehistoric people settled down near water resources around the region. The city is full of major scientific institutes and universities, reflecting the spirit of 1993 Daejeon Expo.
Transportation: Daejeon is connected from/to Incheon Airport by airport limousine and train.
Weather: Average 8°C/-3°C at the end of November or in the beginning of December.
More information at English Wikipedia, Daejeon City, and Korean Tourism Organization.
You can sign up for email notifications acat-info@cern.ch by sending email to acat-loc2021@cern.ch! This list is low traffic and will only get you ACAT conference announcements and general information (for this and future conferences in the ACAT series).
Many people are working together to bring you this conference! The organization page has some details. David Britton is the chair of the International Advisory Committee and Axel Naumann is the chair of the Scientific Program Committee. Soonwook Hwang and Doris Kim are the co-chairs of the Local Organizing Committee.
The World Health Organization has been and is monitoring the development of the pandemic through the regular collection of disease and laboratory data from all member states. Data is collected on the number of cases and death, the age distribution, infections in health care workers, but also on what public health measures are taken and where how many people are vaccinated. This data allows that interventions can be targeted and countermeasures taken. While data is available there have been many challenges to report and clean the data and to do analysis in a timely manner. WHO is collaborating with different groups to enhance the analysis and is looking further to bring more advanced analytic methods forward for the future. Working multidisciplinary will help us understand the pandemic better, which will enable better decisions to be taken to respond.
Over the next years, measurements at the LHC and the HL-LHC will provide us with a wealth of data. The best hope of answering fundamental questions like the nature of dark matter, is to adopt big data techniques in simulations and analyses to extract all relevant information.
On the theory side, LHC physics crucially relies on our ability to simulate events efficiently from first principles. In the coming LHC runs, these simulations will face unprecedented precision requirements to match the experimental accuracy. Innovative ML techniques like generative models can help us overcome limitations from the high dimensionality of the parameter space. Such networks can be employed within established simulation tools or as part of a new framework. Since neural networks can be inverted, they open new avenues in LHC analyses.
This track includes topics that impact how we do physics analysis and research that are related to the enabling technology.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
Problematic I/O pattern is the major cause of low efficiency HEP jobs. When the computing cluster is partially occupied by jobs with problematical I/O patterns, the overall CPU efficiency will dramatically drop down. In a cluster with thousands of users, locating the source of an anomalous workload is not an easy task. Automatic anomaly detection of I/O behavior can largely alleviate the impact of these situations and reduce the manpower spent problem diagnoses. A job’s I/O behavior mainly includes tens of metadata operations such as open, close, getattr etc., and tens of data operations such as read, write etc. Manually setting a problematic threshold for operation cannot adapt to the diversity and variability of the cluster.
This paper provides a data driven method to solve this problem. First, we collect I/O behavior information of each job from the job statistics monitoring file of Lustre file system through collectD and insert them into an Elasticsearch database. Then we search, aggregate and assemble these items into data samples which can be used by machine learning algorithms. After that, we can train unsupervised models with data samples per week and per day. Finally, we can make almost real time anomaly detection by the anomalous score generated given by pre trained models for a new data sample. Currently, the unsupervised model we used is Isolation Forest, which is a very efficient and scalable algorithm for point anomaly detection in a high dimension space. In the future, we can leverage more comprehensive models such as LSTM to make detections of job behavior as a sequence of I/O pattern in its life time.
These tool has been deployed in our production system. It collects tens of thousands of samples per day, hundreds of thousands samples per week, which makes sufficient statistics basis to build an isolation forest. Visualization and sorting tools on web page is also provided to facilitate problem diagnosis and validation of our idea.
Future HEP experiments will have ever higher read-out rate. It is then essential to explore new hardware paradigms for large scale computations. In this work we consider the Optical Processing Unit (OPU) from LightOn, which is an optical device allowing to compute in a fast analog way the multiplication of an input vector of size 1 million by a 1 million x 1 million fixed random matrix, which allows efficient dimension reduction (see e.g. arXiv:2107.11814 ).
This new computing paradigm has so far not been explored in HEP.
We've used the dataset provided by authors of arXiv:1807.00083 who had explored the use of Convolutional Neural network to classify LHC proton proton collision events directly from the raw calorimeter information (without cluster or jet building).
We've proceeded on setting up a pipeline to map the calorimeter information to the OPU input and train on the OPU correlated output random features with a linear model. We show that the pipeline provides classifying power both directly, or as a complement to more traditional jet information.
We present a decisive milestone in the challenging event reconstruction of the CMS High Granularity Calorimeter (HGCAL): the deployment to the official CMS software of the GPU version of the clustering algorithm (CLUE). The direct GPU linkage of CLUE to the preceding energy deposits calibration step is thus made possible, avoiding data transfers between host and device, further extending the heterogeneous chain of HGCAL's reconstruction framework. In addition to various changes and improvements in the management of device memory, new recursive device kernels are added. The latter efficiently navigate through the hit-level information provided by CLUE, calculating the position and energy of the clusters, which are then stored in a cluster-level condensed format. Data conversions from GPU to CPU are included, using structures of arrays, to facilitate the validation of the algorithms and increase the flexibility of the reconstruction chain. For the first time in HGCAL, conditions data (in this case the position of detector elements) are deployed to and filled directly in the GPU and made available on demand at any stage of the heterogeneous reconstruction. This is achieved via a new geometry ordering scheme where a strong correlation between physical and memory locations is present. This scheme is successfully tested with the GPU CLUE version here reported, but is expected to come with a broad range of applicability, and be used by future heterogeneous developments in CMS. Finally, the performance of the combined calibration and clustering algorithms on GPU is assessed and compared to its CPU counterpart.
The Worldwide LHC Computing Grid (WLCG) is the infrastructure enabling the storage and pro-cessing of the large amount of data generated by the LHC experiments, and in particular the ALICE experiment among them. With the foreseen increase in the computing requirements of the future HighLuminosity LHC experiments, a data placement strategy which increases the efficiency of the WLCG computing infrastructure becomes extremely relevant for the scientific success of the LHC scientificprogramme. Currently, the data placement at the ALICE Grid computing sites is optimised via heuristic algorithms. Optimisation of the data storage could yield substantial benefits in terms of efficiency and time-to-result. This has however proven to be arduous due to the complexity of the problem. In this work we propose a modelisation of the behaviour of the system via principal component analysis, time series analysis and deep learning, starting from the detailed data collected by the MonALISA monitoring system. We show that it is possible to analyse and model the throughput of the ALICE Grid to a level that has not been possible before. In particular we compare the performance of different deep learning architectures based on recurrent neural networks. Analyzing about six weeks of ALICE Grid I/O, the trend of the throughput is successfully predicted with a mean relative error of ~4%, while the prediction of the throughput itself is achieved with an accuracy of ~5%.
An accurate prediction of the MonALISA system behavior can lead to a reduction of the time to answer client queries since the in-memory learned model could be used instead of querying the database.
Programmers using the C++ programming language are increasingly taught to manage memory implicitly through containers provided by the C++ standard library. However, many heterogeneous programming platforms require explicit allocation and deallocation of memory, which is often discouraged in “best practice” C++ programming, and this discrepancy in memory management strategies can be daunting for C++ developers who are not already familiar with heterogeneous programming. The C++17 standard introduces the concept of memory resources, which allows the user to control how standard library containers allocate memory. By default, memory resources only support host memory, but their use can be extended to include device memory to relieve the aforementioned issue of ergonomics. In this talk, we present vecmem, a library of memory resources for heterogeneous computing, which allows efficient and user-friendly allocation of memory on CUDA, HIP, and SYCL devices through standard C++ containers. In addition, it provides a number of STL-like classes, which present a familiar programming interface in device code for accessing, and even modifying, the contents of various types of containers. We investigate the design and use cases of such a library, the potential performance gains over naive memory allocation, and the limitations of this memory allocation model.
In the domain of high-energy physics (HEP), query languages in general and SQL in particular have found limited acceptance. This is surprising since HEP data analysis matches the SQL model well: the data is fully structured and queried using mostly standard operators. To gain insights on why this is the case, we perform a comprehensive analysis of six diverse, general-purpose data processing platforms and compare them with ROOT's RDataFrame interface executing the Analysis Description Languages (ADL) benchmark. We identify 16 language features that are useful in implementing typical query patterns found in HEP analyses, categorize them in terms of how essential they are, and analyze how well the different query interfaces implement them. The result of the evaluation is an interesting and rather complex picture of existing solutions: Their query languages vary greatly in how natural and concise HEP query patterns can be expressed but the best-suited languages arguably allow for more elegant query formulations than RDataFrames. At the same time, most of them are also between one and two orders of magnitude slower than that system when tested on large data sets. These observations suggest that, while database systems and their query languages are in principle viable tools for HEP, significant performance improvements are necessary to make them relevant in practice.
There are as many different algorithms and methods as their are physicists. Obviously, we can't list them all here, but here are some broad outlines of techniques that fit into this category. Of course, new and novel categories are part of what this conference is looking for.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
We show how to adapt and deploy anomaly detection algorithms based on deep autoencoders, for the unsupervised detection of new physics signatures in the extremely challenging environment of a real-time event selection system at the Large Hadron Collider (LHC). We demonstrate that new physics signatures can be enhanced by three orders of magnitude, while staying within the strict latency and resource constraints of a typical LHC event filtering system. This would allow for collecting datasets potentially enriched with high-purity contributions from new physics processes. Through per-layer, highly parallel implementations of network layers, support for autoencoder-specific losses on FPGAs and latent space based inference, we demonstrate that anomaly detection can be performed in as little as 80 ns using less than 3% of the logic resources in the Xilinx Virtex VU9P FPGA. Opening the way to real-life applications of this idea during the next data-taking campaign of the LHC.
We present a machine-learning based strategy to detect data departures from a given reference model, with no prior bias on the nature of the new physics responsible for the discrepancy. The main idea behind this method is to build the likelihood-ratio hypothesis test by directly translating the problem of maximizing a likelihood-ratio into the minimization of a loss function. A neural network compares observations with an auxiliary set of reference-distributed events, possibly obtained with a Monte Carlo event generator. The virtues of neural networks as unbiased function approximants make them particularly suited for this task. The algorithm returns a p-value, which measures the compatibility of the reference model with the data. It also identifies the most discrepant phase-space region of the data set, to be selected for further investigation.
The most interesting potential applications are new physics searches in high energy physics, for which our work provides an end-to-end signal-model-independent analysis strategy.
Besides that, our approach could also be used to compare the theoretical predictions of different Monte Carlo event generators, or for data validation algorithms.
In this talk, after outlining the conceptual foundations of the algorithm [1], we explain how to apply it to a multivariate problem [2] and how to extend it to deal with uncertainties on the reference model predictions by studying the impact of two typical sources of experimental uncertainties in a two-body final state analysis at the LHC.
[1] https://link.aps.org/doi/10.1103/PhysRevD.99.015014
[2] https://doi.org/10.1140/epjc/s10052-021-08853-y
Photometric data-driven classification of supernovae is one of the fundamental problems in astronomy. Recent studies have demonstrated the superior quality of solutions based on various machine learning models. These models learn to classify supernova types using their light curves as inputs. Preprocessing of these curves is a crucial step that significantly affects the final quality. In this talk, we study the application of shallow neural networks, bayesian neural networks, and normalizing flows to approximate observations for a single light curve. We use these approximations as inputs for supernovae classification models and demonstrate that the proposed methods outperform the state-of-the-art based on gaussian processes applying to the PLAsTiCC synthetic dataset. As an additional experiment, we perform intensity peak estimation. We use different approximations of light curves and fit a convolutional neural network to predict an intensity peak position. Results demonstrate that proposed algorithms help to improve the accuracy of the peak estimation compared to the gaussian processes model. We also apply the algorithms to the Bright Transient Survey ZTF light curves. Shallow neural networks demonstrate similar results as gaussian processes showing factor 5 increased speed. Normalizing Flows exceeds gaussian processes in terms of approximation quality as well.
Over the last ten years, the popularity of Machine Learning (ML) has grown exponentially in all scientific fields, included particle physics. Industry has also developed new powerful tools that, imported into academia, could revolutionise research. One recent industry development that has not yet come to the attention of the particle physics community is Collaborative Learning (CL), a framework which allows training the same ML model with different datasets. This work explores the potential of CL, testing the library Colearn with neutrino physics simulation. Colearn, developed by the British Cambridge-based firm Fetch.AI, enables decentralised machine learning tasks. Being a blockchain-mediated CL system, it allows multiple stakeholders to build a shared ML model without needing to rely on a central authority. A generic Liquid Argon Time-Projection Chamber (LArTPC) has been simulated and images produced by fictitious neutrino interactions have been used to produce several datasets. These datasets, called learners, participated successfully in training a Deep Learning (DL) keras model using blockchain technologies in a decentralised way. This test explores the feasability of training a single ML model using different simulation datasets coming from different research groups. In this work, we also discuss a framework that instead makes different ML models competing against each other on the same dataset. The final goal is then to train the most performant ML model across the entire scientific community for a given experiment, either using all of the datasets available or selecting the model which performs best among every model developed in the community.
The Self-Organizing-Map (SOM) is a widely used neural
net for data analysis, dimension reduction and
clustering. It has yet to find use in high energy
particle physics. This paper discusses two
applications of SOM in particle physics. First, we were
able to obtain high separation of rare processes in
regions of the dimensionally reduced representation.
Second, we obtained Monte Carlo scale factors by fitting
the dimensionally reduced representation.
Analysis and training were performed on the data of the
ATLAS Machine Learning challenge and on Open
Data.
We investigate supervised and unsupervised quantum machine learning algorithms in the context of typical data analyses at the LHC. To deal with constraints on the problem size, dictated by limitations on the quantum hardware, we concatenate the quantum algorithms to the encoder of a classic autoencoder, used for dimensional reduction. We show results for a quantum classifier and a quantum anomaly detection algorithm, comparing performance to corresponding classic algorithms.
This track focuses on computing techniques and algorithms used in the theoretical side of physics research.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
Amplitude level evolution has become a new theoretical paradigm to analyze parton shower algorithms which are at the heart of multi-purpose event generator simulations used for particle collider experiments. It can also be implemented as a numerical algorithm in its own right to perform resummation of non-global observables beyond the leading colour approximation, leading to a new kind of parton cascade. In this talk I will present the computational details and challenges behind the CVolver method and library, and its relation to new Monte Carlo sampling algorithms and a possible future structure of event generators.
In this talk we present a neural network based model to emulate matrix
elements. This model improves on existing methods by taking advantage of the known
factorisation properties of matrix elements to separate out the divergent regions.
In doing so the neural network learns about the factorisation property in singular limits, meaning we can control the behaviour of simulated matrix elements when extrapolating into more singular regions than the ones used for training the neural network. We apply our model to the case of leading-order jet production in $e^{+}e^{−}$ collisions with up to five jets.
Our results show that this model can reproduce the matrix elements with errors below the one-percent level on the phase-space covered during fitting and testing, and a robust extrapolation to the parts of the phase-space where the matrix elements are more singular than seen at the fitting stage. We also demonstrate that the uncertainties associated with the neural network predictions are negligible compared to the statistical Monte Carlo errors.
Talk based on [2107.06625] with code to reproduce the method at https://www.github.com/htruong0/fame
The demand for precision predictions in the field of high energy physics has increased tremendously over the recent years. Its importance is visible in the light of current experimental efforts to test the predictive power of the Standard Model of particle physics (SM) to a never before seen accuracy. Thus, advanced computer software is a key technology to enable phenomenological computations for the needed SM predictions.
In this background we present tapir
: A tool for identification, manipulation and minimization of Feynman diagrams and their corresponding topologies. It is designed to integrate in FORM
-based toolchains which is a common practice in the field. tapir
can be used to reduce the complexity of multi-loop problems with cutting-filters, scalar topology mapping, partial fraction decomposition and alike.
Scattering amplitudes in perturbative quantum field theory exhibit a rich structure of zeros, poles and branch cuts which are best understood in complexified momentum space. It has been recently shown that leveraging this information can significantly simplify both analytical reconstruction and final expressions for the rational coefficients of transcendental functions appearing in phenomenologically-relevant scattering amplitudes. Inspired by these observations, we present a new algorithmic approach to the reconstruction problem based on p-adic numbers and computational algebraic geometry. For the first time, we systematically identify and classify the relevant irreducible varieties in spinor space, and thanks to p-adic numbers -- analogous finite fields, but with a richer structure to their absolute value -- we stably perform numerical evaluations close to these singular surfaces, thus completely avoiding the use of floating-point numbers. Finally, we discuss a GPU-based implementation of dense Gaussian elimination used to perform some of the linear algebra steps involved in the construction of the ansatz and fitting of its coefficients.
In this talk I will present REvolver, a c++ library for renormalization group evolution and automatic flavor matching of the QCD coupling and quark masses, as well as precise conversion between various quark mass renormalization schemes. The library systematically accounts for the renormalization group evolution of low-scale short-distance masses which depend linearly on the renormalization scale and sums logarithmic terms of high and low scales that are missed by the common logarithmic renormalization scale evolution. The library can also be accessed through Mathematica and Python nterfaces and provides renormalization group evolution for complex renormalization scales as well. In the presentation I will demonstrate how the library works in real time.
pySecDec is a tool for Monte Carlo integration of multiloop Feynman integrals (or parametric integrals in general), using the sector decomposition strategy. Its latest release contains two major features: the ability to expand integrals in kinematic limits using expansion by regions approach, and the ability to optimize the integration of weighted sums of integrals maximizing the obtained precision within a constant time. I'd like to present these features in particular (based on arXiv:2108.10807), and the inner workings of pySecDec in general.
Non perturbative QED is used to predict beam backgrounds at the interaction point of colliders, in calculations of Schwinger pair creation and in precision QED tests with ultra-intense lasers.
In order to predict these phenomena, custom built monte carlo event generators based on a suitable non perturbative theory have to be developed. One such suitable theory uses the Furry Interaction Picture, in which a background field is taken into account non perturbatively at Lagrangian level. This theory is precise, but the transition probabilities are in general, complicated. This poses a challenge for the monte carlo which struggles to implement the theory computatively. The monte carlo must in addition taken into acount the behaviour of the background field at every space-time point at which an event is generated. We introduce here just such a monte carlo package, called IPstrong, and the techniques implemented to deal with the specific challenges outlined above.
Multiloop calculations are vital for obtaining high-precision predictions in Standard Model. In particular, such predictions are important for the possibility to discover New Physics which is expected to reveal itself in tiny deviations. The methods of multiloop calculations are rapidly evolving for already a few decades. New algorithms as well as their specific software implementations appear every year. In the present talk I will give my personal view on recent advances in multiloop methods and possible directions of their further development.
Autonomous driving is an extremely hot topic, and the whole automotive industry is now working hard to transition from research to products. Deep learning and the progress of silicon technology are the main enabling factors that boosted the industry interest and are currently pushing the automotive sector towards futuristic self-driving cars. Computer vision is one of the most important sensing technologies thanks to the extremely dense information it can provide as output: object classification and recognition, precise localization, 3D reconstruction are just examples. This presentation addresses the missing piece that will allow future vehicles to make a full and efficient use of computer vision.
The unprecedented volume of data and Monte Carlo simulations at the HL-LHC will pose increasing challenges for data analysis both in terms of computing resource requirements as well as "time to insight". I will discuss the evolution and current state of analysis data formats, software, infrastructure and workflows at the LHC, and the directions being taken towards fast, efficient, and effective physics analysis at the HL-LHC.
This track includes topics that impact how we do physics analysis and research that are related to the enabling technology.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
From 2022 onward, the upgraded LHCb experiment will use a triggerless readout system collecting data at an event rate of 30 MHz. A software-only High Level Trigger will enable unprecedented flexibility for trigger selections. During the first stage (HLT1), a sub-set of the full offline track reconstruction for charged particles is run to select particles of interest based on single or two-track selections. This includes decoding the raw data, clustering of hits, pattern recognition, as well as track fitting, ghost track rejection with machine learning techniques and finally event selections. HLT1 will be processed entirely on O(100) GPUs. We will present the integration of the GPU HLT1 into LHCb's data acquisition system in preparation for data taking in 2022.
At the start of the upcoming LHC Run-3, CMS will deploy a heterogeneous High Level Trigger farm composed of x86 CPUs and NVIDIA GPUs. In order to guarantee that the HLT can run on machines without any GPU accelerators - for example as part of the large scale Monte Carlo production running on the grid, or when individual developers need to optimise specific triggers - the HLT reconstruction has been implemented both for NVIDIA GPUs and for traditional CPUs. This contribution will describe how the CMS software used online and offline (CMSSW) can transparently switch between the two implementations, and compare their performance on GPUs and CPUs from different architectures, vendors and generations.
After the Phase II Upgrade of the LHC, expected for the period between 2025-26, the average
number of collisions per bunch crossing at the LHC will increase from the Run-2 average value
of 36 to a maximum of 200 pile-up proton-proton interactions per bunch crossing. The ATLAS
detector will also undergo a major upgrade programme to be able to operate it in such a harsh
conditions with the same or better performance than up now.
The ATLAS Trigger system, responsible for the online processing and filtering of the
collisions happening at the LHC, will have to analyse the 40 MHz collisions rate, selecting
about 10 kHz for offline analysis. The trigger is a tiered system, consisting in a first selection
level made of custom hardware processors, that use reduced granularity from the calorimeters
and muon detectors, followed by a High Level Trigger implemented in software, running on a
farm of commodity computing and benefitting from the full detector granularity and improved
calibrations. The high pile-up conditions expected after the upgrades, will be very challenging
for the Trigger system, that will have to cope with increased rates while keeping the same
efficiency for the interesting physics processes. This will require the use of advanced, more time
consuming algorithms.
The ATLAS Collaboration is currently studying the use of general purpose Graphical
Processing Units (GPUs) as hardware accelerators in trigger and offline reconstruction, as a
mean to deal with the increasing processing times demands in view of the upgrades. Since
the first GPU Trigger Prototype [1], that demonstrated the potential gains (and limitations)
of GPUs for the trigger system, a new implementation of the calorimeter TopoCluster [2]
reconstruction algorithm for GPU has been done, benefitting from the new multithreaded
AthenaMT framework [3] and a simplified architecture.
This communication will describe the new implementation of the calorimeter clustering
algorithm, and will present detailed performance studies, including the resulting speed-up of
the algorithm processing times.
The z-vertex track trigger estimates the collision origin in the Belle II experiment using neural networks to reduce the background. The main part is a pre-trained multilayer perceptron. The task of this perceptron is to estimate the z-vertex of the collision to suppress background from outside the interaction point. For this, a low latency real-time FPGA implementation is needed. We present an overview of the architecture and the FPGA implementation of the neuronal network and the preprocessing. We also show the hardware preprocessing handling of missing input data with specially trained neuronal networks. For this, we will show the results of the z-vertex estimation and the latency for the implementation in the Belle II trigger system. Plans for a major update, for the preprocessing, with a 3D Hough transformation processing step is ongoing and will be presented.
Several online and offline applications in high-energy physics have benefitted from running on graphics processing units (GPUs), taking advantage of their processing model. To date, however, general HEP particle transport simulation is not one of them, due to difficulties in mapping the complexity of its components and workflow to the GPU’s massive parallelism features. Deep code stacks, with polymorphism, low branch predictability, incoherent memory accesses, and the use of stateful global managers are significant obstacles preventing porting Geant4 to GPUs. HEP computing will need to exploit more and more heterogeneous resources in the future, and our current inability to use GPU cards for detailed simulation of collider experiments limits the total performance available and raises costs, motivating R&D in this area.
The AdePT project is one of the R&D initiatives tackling this limitation and exploring GPUs as potential accelerators for offloading some part of the CPU simulation workload. We started one year ago with the ambitious goal of demonstrating a complete workflow working on GPU, having all the simulation stepping components: complete physics models describing the electromagnetic processes, magnetic field propagation in detector geometry, and code producing user hits
data transferred from the GPU back to the host. The project is the first to create a full prototype of a realistic electron, positron, and gamma EM shower simulation on GPU implemented as either a standalone application or as an extension of the standard Geant4 CPU workflow.
Most of the original goals have already been achieved, and the prototype provides a platform to explore many optimisations and different approaches. We will present the most recent results and initial conclusions of our work, including a performance study comparing standalone and hybrid workflows on the CPU and GPU. We will describe the main features and components of the demonstrators developed in the AdePT project, the optimization process, and our preliminary understanding of the usability of GPUs for full simulation in HEP applications.
Automatic Differentiation (AD) is instrumental for science and industry. It is a tool to evaluate the derivative of a function specified through a computer program. The range of AD application domain spans from Machine Learning to Robotics to High Energy Physics. Computing gradients with the help of AD is guaranteed to be more precise than the numerical alternative and have at most a constant factor (4) more arithmetical operations compared to the original function. Moreover, AD applications to domain problems typically are computationally bound. They are often limited by the computational requirements of high-dimensional transformation parameters and thus can greatly benefit from parallel implementations on graphics processing units (GPUs).
Clad aims to enable differentiable analysis for C/C++ and CUDA and is a compiler-assisted AD tool available both as a compiler extension and in ROOT. Moreover, Clad works as a compiler plugin extending the Clang compiler; as a plugin extending the interactive interpreter Cling; and as a Jupyter kernel extension based on xeus-cling.
In this talk, we demonstrate the advantages of parallel gradient computations on graphics processing units (GPUs) with Clad. We explain how to bring forth a new layer of optimisation and a proportional speed up by extending the usage of Clad for CUDA. The gradients of well-behaved C++ functions can be automatically executed on a GPU. Thus, across the spectrum of fields, researchers can reuse their existing models and have workloads scheduled on parallel processors without the need to optimize their computational kernels. The library can be easily integrated into existing frameworks or used interactively, and provides optimal performance. Furthermore, we demonstrate the achieved application performance improvements, including (~10x) in ROOT histogram fitting and corresponding performance gains from offloading to GPUs.
We present results from a stand-alone simulation of electron single coulomb scattering as implemented completely on an FPGA architecture and compared with an identical simulation on a standard CPU. FPGA architectures offer unprecedented speed-up capability for Monte Carlo simulations, however with the caveats of lengthy development cycles and resource limitation particularly in terms of on-chip memory and DSP blocks. As a proof of principle of acceleration on an FPGA we chose a single scattering process of electrons in water at an energy of 6 MeV. The initial code-base was implemented in c++ and optimised for CPU processing. To measure the potential performance gains of FPGAs compared to modern multi-core CPUs we computed 100M histories of a 6 MeV electron interacting in water. The FPGA bit-stream is implemented using MaxCompiler 2021.1 and Vivado 2019.2. MaxCompiler is a High-Level Synthesis (HLS) language that facilitates implementation between CPU and FPGAs; it greatly reduces the development time but does not achieve the same performance as manually optimised VHDL. We did not perform any hardware specific optimisation. We also limited the clock frequency to only 200 MHz, which is easily achievable by any HLS implementation on a modern FPGA. The same arithmetic precision was applied to the FPGA as the CPU implementation. The system configuration comprises an AMD Ryzen 5900x 12-cores CPU running at 3.7 GHz and boosting up to 4.8GHz with a Xilinx's Alveo U200 Data Center accelerator card. The Alveo U200 incorporates a VU9P FPGA device, with a capacity of 1,182,240 LUTs, 2,364,480 FFs, 6,840 DSPs, 4,320 BRAMs and 960 URAMs. The results shows that the FPGA implementation is over 110 times faster than an optimised parallel implementation running on 12-cores and over 270x faster than a sequential single core implementation. For today's market prices, this shows a cost equivalent speed-up of more than 10. The results on both architectures were statistically equivalent. The successful implementation and measured acceleration is very encouraging for future exploits of more generic Monte Carlo simulation on FPGAs for High Energy Physics applications.
There are as many different algorithms and methods as their are physicists. Obviously, we can't list them all here, but here are some broad outlines of techniques that fit into this category. Of course, new and novel categories are part of what this conference is looking for.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
High energy physics experiments relies on Monte Carlo simulation to accurately model their detector response. Most of the time dominated by shower simulation in the calorimeter, the detector response modelling is time consuming and CPU intensive especially with the upcoming High Luminosity LHC upgrade. Several research directions investigated the use of Machine Learning based models to accelerate particular detector response simulation. This results in a specifically tuned simulation and generally these models require a large amount of data for training. Meta learning has emerged recently as fast learning algorithm using a small training dataset. In this project, we build a meta-learning model that “learns how to learn” to generate showers using a first-order gradient based algorithm. This model is trained on multiple detector geometries and can rapidlly adapt to a new geometry using few training samples.
The increasing luminosities of future data taking at Large Hadron Collider and next generation collider experiments require an unprecedented amount of simulated events to be produced. Such large scale productions demand a significant amount of valuable computing resources. This brings a demand to use new approaches to event generation and simulation of detector responses. In this talk, we discuss the application of generative adversarial networks (GANs) to the simulation of the LHCb experiment events. We emphasize main pitfalls in the application of GANs and study the systematic effects in detail. The presented results are based on the Geant4 simulation of the LHCb Cherenkov detector.
Generative models (GM) are powerful tools to help validate theories by reducing the computation time of Monte Carlo (MC) simulations. GMs can learn expensive MC calculations and generalize to similar situations. In this work, we propose comparing a classical generative adversarial network (GAN) approach with a Born machine, both in his discrete (QCBM) and continuous (CVBM) form while addressing their strength and limitations, to generate muon force carrier (MFC) events. The former uses a neural network as a discriminator to train the generator, while the latter takes advantage of the probabilistic nature of quantum mechanics to generate samples. We consider a muon fixed\hyp target collison from the ForwArd Search ExpeRiment (FASER) at the large hadron collider (LHC), with the ATLAS calorimeter as the target. The independent muon measurements performed by the inner detector (ID) and muon system (MS) can help to observe new force carriers coupled to muons, which are usually not detected. We concentrate on muons coming from W and Z bosons decays. MFCs could potentially be part of dark matter (DM), making them interesting for physic searches beyond the standard model.
AtlFast3 is the next generation of high precision fast simulation in ATLAS. It is being deployed by the collaboration and will replace AtlFastII, the fast simulation tool that was successfully used until now. AtlFast3 combines two Fast Calorimeter Simulations tools; a parameterization-based approach and a machine-learning based tool exploiting Generative Adversarial Networks (GANs). AtlFast3 provides improved agreement with Geant4 for objects used in analyses, with a focus on those that were poorly modelled in AtlFastII. In particular, the simulation of jets of particles reconstructed with large radii and the detailed description of their substructure, are significantly improved in AtlFast3. The modelling and performance are evaluated on events produced at 13 TeV centre-of-mass energy in the Run-2 data-taking conditions.
SND@LHC is a newly approved detector under construction at the LHC, aimed at studying the interactions of neutrinos of all flavours produced by proton-proton collisions at the LHC. The energy range under study, few hundreds MeVs up to about 5 TeVs, is currently unexplored. In particular, electron neutrino and tau neutrino cross sections are unknown in that energy range, whereas muon neutrino data stops at a few hundreds GeVs. The SND@LHC detector will allow to fill in these gaps for all three flavours, enabling crucial tests of the SM. The detector is also ideally suited to look for scattering signatures of hypothetical light Dark Matter particles.
The neutrino interaction target is made of alternating layers of tungsten plates, nuclear emulsion films and scintillating fiber (SciFi) trackers. As such, it provides micrometric accuracy for track reconstruction when the emulsion films are taken out for development, every few months. However, the presence of the SciFi layers allow for real-time event analysis.
This talks presents for the first time a new approach to calorimetry with a tracking detector. Conventional tracking methods are not useful in the context of high-multiplicity particle showers, as are expected in neutrino scattering events. Machine learning techniques will therefore be employed in order to analyse data from the SciFi tracker in real time, providing a measurement of the neutrino energy and flavour tagging. Preliminary studies prove that the SciFi planes do behave as the active layers of a sampling calorimeter. Various techniques based on Neural Networks (NN), including Bayesan and Convolutional NNs, are being developed, allowing to overcome technical challenges such as the presence of ghost hits, and to supplement calorimetric information with topological event reconstruction, a first-time attempt in the field.
The idea reported in this talk represents a general breakthrough for the reconstruction of complex events with an electronic detector, a crucial asset in view of operating a particle detector at a high luminosity collider or beam dump, and much more so for a neutrino experiment requiring flavour tagging. This talk will give an overview of the results already achieved and of the short term prospects. These results had never been publicly presented before.
The LUXE experiment (LASER Und XFEL Experiment) is a new experiment in planning at DESY Hamburg that will study Quantum Electrodynamics (QED) at the strong-field frontier. In this regime, QED is non-perturbative. This manifests itself in the creation of physical electron-positron pairs from the QED vacuum. LUXE intends to measure the positron production rate in this unprecedented regime by using, among others, a silicon tracking detector. The large number of expected positrons (up to O($10^4$)) traversing the sensitive detector layers results in an extremely challenging combinatorial problem which can become computationally very hard for classical computers. After an overview of the LUXE experimental setup and simulation, this talk will present a preliminary study to explore the potential of quantum computers to solve this problem and to reconstruct the positron trajectories from the detector energy deposits. The reconstruction problem is formulated in terms of a quadratic unconstrained binary optimisation. Finally the results from the quantum simulations are discussed and compared with traditional classical track reconstruction algorithms.
This track focuses on computing techniques and algorithms used in the theoretical side of physics research.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
We introduce the differentiable simulator MadJax, an implementation of the general purpose matrix element generator Madgraph integrated within the Jax differentiable programming framework in Python. Integration is performed during automated matrix element code generation and subsequently enables automatic differentiation through leading order matrix element calculations. Madjax thus facilitates the incorporation of high energy physics domain knowledge, encapsulated in simulation software, into gradient based learning and optimization pipelines. In this paper we present the MadJax framework as well as several example applications enabled uniquely through the capabilities of differentiable simulation.
We present the software framework underlying the NNPDF4.0 global determination of parton distribution functions (PDFs). The code is released under an open source licence and is accompanied by extensive documentation and examples. The code base is composed by a PDF fitting package, tools to handle experimental data and to efficiently compare it to theoretical predictions, and a versatile analysis framework. In addition to ensuring the reproducibility of the NNPDF4.0 (and subsequent) determination, the public release of the NNPDF fitting framework enables a number of phenomenological applications and the production of PDF fits under user-defined data and theory assumptions.
We present Qibo, a new open-source framework for fast evaluation of quantum circuits and adiabatic evolution which takes full advantage of hardware accelerators, quantum hardware calibration and control, and large codebase of algorithms for applications in HEP and beyond. The growing interest in quantum computing and the recent developments of quantum hardware devices motivates the development of new advanced computational tools focused on performance and usage simplicity. In this work we introduce a new quantum simulation framework that enables developers to delegate all complicated aspects of hardware or platform implementation to the library so they can focus on the problem and quantum algorithms at hand. As example for HEP applications, we show how to use Qibo for the determination of parton distribution functions (PDFs) using DIS, fixed-target DY and LHC data, and the construction of generative models applied to Monte Carlo simulation. We conclude by providing an overview of the variational quantum circuit models included in Qibo.
We compute the coefficients of the perturbative expansions of the plaquette,
and of the self-energy of static sources in the triplet and octet representation,
up to very high orders in perturbation theory. We use numerical sthocastic
perturbation theory and lattice regularization. We explore if the results
obtained comply with expectations from renormalon dominance, and what
they may say for a model independent and nonperturbative determination
of the value of the gluon condensate.
Modern machine learning methods offer great potential for increasing the efficiency of Monte Carlo event generators. We present the latest developments in the context of the event generation framework SHERPA. These include phase space sampling using normalizing flows and a new unweighting procedure based on neural network surrogates for the full matrix elements. We discuss corresponding general construction criteria and show examples of efficiency gains for relevant LHC production processes.
Normalizing Flows (NFs) are emerging as a powerful brand of generative models, as they not only allow for efficient sampling, but also deliver density estimations by construction. They are of great potential usage in High Energy Physics (HEP), where we unavoidably deal with complex high dimensional data and probability distributions are everyday’s meal. However, in order to fully leverage the potential of NFs it is crucial to explore their robustness as the dimensionality of our data increases. Thus, in this talk, we discuss the performance of some of the most popular types of NFs on the market, on several example data sets with escalating number of dimensions.
Human hearing has a very amazing ability that even advanced technology cannot imitate. The difference in energy between the smallest and loudest audible sounds is about a trillion times. The frequency resolution is also excellent, so the ear can distinguish a frequency difference of about 4 Hz. What is more surprising is that it can be heard even where there is a louder noise than the sound of speech, and it can detect the direction of the sound and focus on listening to one of several people's speech. Many efforts have been made to understand this principle of hearing, but it is not yet fully understood. In this talk, we will show that if we create an artificial neural network that mimics human hearing and train a machine just like a human learns, it can surprisingly have the characteristics of human hearing.
In this talk, I will discuss the impacts of what has been termed a growing culture of speed and hypercompetition in the academic sciences. Drawing on qualitative social sciences research in the life sciences, I will discuss how acceleration and hypercompetition impact epistemic diversity in science, i.e. the range of research topics researchers consider they can address, as well as human diversity, i.e. diversity related to gender, social class and other social factors. With the audience, I want to discuss these findings in relation to their own experiences of working in their fields of research as well as concurrent initiatives to address the negative impacts of acceleration and hypercompetition on individual, institutional and policy levels.
The affinity between statistical physics and machine learning has a long history, I will describe the main lines of this long-lasting friendship in the context of current theoretical challenges and open questions about deep learning. Theoretical physics often proceeds in terms of solvable synthetic models, I will describe the related line of work on solvable models of simple feed-forward neural networks. I will highlight a path forward to capture the subtle interplay between the structure of the data, the architecture of the network, and the learning algorithm.
This track includes topics that impact how we do physics analysis and research that are related to the enabling technology.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
The Super Tau Charm Facility (STCF) is a high-luminosity electron–positron
collider proposed in China, for the study of charm and tau physics. The Offline Software of Super Tau Charm Facility (OSCAR) is designed and developed
based on SNiPER, a lightweight common framework for HEP experiments. Several state-of-art software and tools in the HEP community are adopted, such as
the Detector Description Toolkit (DD4hep) for the consistent detector-geometry
description, the plain-old-data I/O (PODIO) for the efficient implementation of
the event data model, etc. This talk will focus on the design and implementation of OSCAR, particularly the way to integrate Geant4, DD4hep and PODIO
into SNiPER to provide a unified computing environment and platform for
detector simulation, reconstruction and visualization. Now OSCAR is used to
facilitate design of the STCF detector, conduct detector performance study as
well as the physics potential study. OSCAR also provides a potential solution for other lightweight HEP experiments.
The JUNO experiment is being built mainly to determine the neutrino mass hierarchy by detecting neutrinos generated in the Yangjiang and Taishan nuclear plants in southern China. The detector will record 2 PB raw data every year, but each day it can only collect about 60 neutrino events scattered among huge background events. Selection of extremely sparse neutrino events brings a big challenge to offline data analysis. A typical neutrino physics event normally spans across a number of consecutive readout events, flagged by a fast positron signal followed by a slow neutron signal within a varying-size time window. To facilitate this analysis, a two-step data processing scheme has been proposed. In the first step (called data preparation), the event index data is produced and skimmed, which only contains information of minimum physics quantities of events as well as their addresses in the original reconstructed data file. In the second step (called time correlation analysis), event index data is further selected with stricter criteria. And then, for each selected event, the time correlation analysis is performed by reading all associated events within a pre-defined time window from the original data file according to the selected event’s address and timestamp.
Firstly, this contribution will introduce the design of the above data processing scheme and then focus on the multi-threaded implementation of time correlation analysis based on the Intel Threading Building Block (TBB) in the SNiPER framework. Secondly, this contribution will describe the implementation of distributed analysis using MPI in which the time correlation analysis task is divided into subtasks running on multiple computing nodes. At last, this contribution will present the detailed performance measurements made on a multiple-node test bed. By using both skimming and indexing techniques, the total amount of data read by time correlation analysis is reduced significantly. So that the processing time could be reduced by two orders of magnitude.
The GeoModel toolkit is an open-source suite of standalone tools that empowers the user with lightweight tools to describe, visualize, test, and debug detector descriptions and geometries for HEP standalone studies and experiments. GeoModel has been designed with independence and responsiveness in mind and offers a development environment free of other large HEP tools and frameworks, and with a very quick development cycle. With very few and lightweight dependencies, GeoModel is easy to install on all systems, in a modular way; and pre-compiled binaries are provided for the major platforms, for a quick and easy installation. Coded entirely in C++, GeoModel offers the user tools to describe geometries inside C++ code or in external XML files, create persistent representation with a low disk footprint and interactively visualize and inspect the geometry in a 3D view. It also offers a plugin mechanism and an optional Geant4 application to simulate the described geometry in a standalone environment. In this contribution, we describe all the available tools, with a focus on the latest additions, which provide users with more visualization, debug, and simulation tools.
A detailed geometry description is essential to any high quality track reconstruction application. In current C++ based track reconstruction software libraries this is often achieved by an object oriented, polymorphic geometry description that implements different shapes and objects by extending a common base class. Such a design, however, has been shown to be problematic when attempting to adapt these application to run on heterogenous computing hardware, particularly on hardware accelerators. We present detray, a compile-time polymorphic and yet accurate track reconstruction geometry description which is part of the ACTS parallelization R&D efforts. detray is built as a indexed based geometry description with dedicated shallow memory layout, and uses variadic template programming to allow custom shapes and intersection algorithms rather than virtual inheritance. It implements the ACTS navigation model of boundary portals and purely surface based entities and is designed to serve as a potential geometry and navigation back-end to ACTS. detray is compatible with the vecmem memory management library and thus can be instantiated as a geometry model in host and device memory, respectively. We present the concepts, a propagation demonstrator using the Open Data Detector and discuss its portability and usage in device applications.
The performance of I/O intensive applications is largely determined by the organization of data and the associated insertion/extraction techniques. In this paper we present the design and implementation of an application that is targeted at managing data received (up to ~150 Gb/s payload throughput) into host DRAM, buffering data for several seconds, matched with the DRAM size, before being dropped. All data are validated, processed and indexed. The features extracted from the processing are streamed out to subscribers over the network; in addition, while data resides in the buffer, about 0.1 ‰ of them are served to remote clients upon request. Last but not least, the application must be able to locally persist data at full input speed when instructed to do so.
The characteristics of the incoming data stream (fixed or variable rate, fixed or variable payload size) heavily influences the choice of implementation of the buffer management system. The application design promotes the separation of interfaces (concepts) and application oriented specializations (models) that makes it possible to generalize most of the workflows and only requires minimal effort to integrate new data sources.
After the description of the application design, we will present the hardware platform used for validation and benchmarking of the software, and the performance results obtained.
The High Luminosity upgrade to the LHC, which aims for a ten-fold increase in the luminosity of proton-proton collisions at an energy of 14 TeV, is expected to start operation in 2028/29, and will deliver an unprecedented volume of scientific data at the multi-exabyte scale. This amount of data has to be stored and the corresponding storage system must ensure fast and reliable data delivery for processing by scientific groups distributed all over the world. The present LHC computing and data management model will not be able to provide the required infrastructure growth even taking into account the expected hardware technology evolution. To address this challenge, the Data Carousel R&D project was launched by the ATLAS experiment in the fall of 2018. By Data Carousel we mean on-demand reading of selected data from tape without pre-staging. Data Carousel uses a sliding window buffer whose size can be tuned to suit available resources and production requirements. The Data Carousel in ATLAS is the orchestration between the workflow management systems, the distributed data management and the tape services. We successfully and quickly passed the R&D project phases involving FTS, dCache, CTA, Rucio, PanDA/JEDI, ATLAS Computing Operations and the WLCG centers. Our current goal is to simultaneously run major ATLAS production workflows in Data Carousel mode with respect to dynamic computing shares and sliding window size. We are also working on tape throughput estimation, in anticipation for HL-LHC. Joint-tape throughput tests with other LHC experiments have also been conducted.
Data Carousel technology may be applicable to other scientific communities, such as SKA, DUNE, Vera Rubin Observatory, BELLE II, and NICA to manage large-scale data volumes between different QoS storage elements. State-of-the-art data and workflow management technologies are under active development and their status will be presented, as well as the ATLAS data carousel plans.
The ALICE experiment at the CERN LHC (Large Hadron Collider) is undertaking a major upgrade during the LHC Long Shutdown 2 in 2019-2021, which includes a new computing system called O2 (Online-Offline). The raw data input from the ALICE detectors will increase a hundredfold, up to 3.5 TB/s. By reconstructing the data online, it will be possible to compress the data stream down to 100 GB/s before storing it permanently.
The O2 software is a message-passing system. It will run on approximately 500 computing nodes performing reconstruction, compression, calibration and quality control of the received data stream. As a direct consequence of having a distributed computing system, locally generated data might be incomplete and could require merging to obtain complete results.
This paper presents the O2 Mergers, the software designed to match and combine partial data into complete objects synchronously to data taking. Based on a detailed study and results of extensive benchmarks, a qualitative and quantitative comparison of different merging strategies considered to reach the final design and implementation of the software is discussed.
There are as many different algorithms and methods as their are physicists. Obviously, we can't list them all here, but here are some broad outlines of techniques that fit into this category. Of course, new and novel categories are part of what this conference is looking for.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
To sustain the harsher conditions of the high-luminosity LHC, the CMS collaboration is designing a novel endcap calorimeter system. The new calorimeter will predominantly use silicon sensors to achieve sufficient radiation tolerance and will maintain highly-granular information in the readout to help mitigate the effects of pileup. In regions characterized by lower radiation levels, small scintillator tiles with individual on-tile SiPM readout are employed. A unique reconstruction framework (TICL: The Iterative CLustering) is being developed within the CMS Software CMSSW to fully exploit the granularity and other significant detector features, such as particle identification and precision timing, with a view to mitigate pileup in the very dense environment of HL-LHC. The TICL framework has been thought with heterogeneous computing in mind: the algorithms and their data structures are designed to be executed on GPUs. In addition, fast and geometry agnostic data structures have been designed to provide fast navigation and searching capabilities. Seeding capabilities (also exploiting information coming from other detectors) dynamic cluster asking, energy regression and particle identification are the main components of the framework. To allow for maximal flexibility, TICL allows composition of different combinations of modules that can be chained together in an iterative fashion. A stable version of TICL has been already integrated in CMSSW and used for the CMS Phase-2 reconstruction in the context of the Phase-2 Upgrade of the CMS Data Acquisition and High Level Trigger Technical Design Report.
In the European Center of Excellence in Exascale Computing "Research on AI- and Simulation-Based Engineering at Exascale" (CoE RAISE), researchers from science and industry develop novel, scalable Artificial Intelligence technologies towards Exascale. In this work, we leverage HPC resources to perform large scale hyperparameter optimization using distributed training on multiple compute nodes, each with multiple GPUs. This is part of CoE RAISE’s work on data-driven use-cases towards Exascale which leverages AI- and HPC cross-methods developed within the project.
Hyperparameter optimization of deep learning-based AI models is often compute resource intensive, partly due to the high cost of training a single hyperparameter configuration to completion, and partly because of the infinite set of possible hyperparameter combinations to evaluate. There is therefore a need for large scale, parallelizable and resource efficient hyperparameter search algorithms. We benchmark and compare different search algorithms for advanced hyperparameter optimization. The evaluated search algorithms, including Random Search, Hyperband and ASHA, are tested and compared in terms of both final accuracy and accuracy per compute resources spent.
As an example use-case, a graph neural network (GNN) model known as MLPF, which has been developed for the task of Machine Learned Particle-Flow reconstruction, acts as the base model for which hyperparameter optimization is performed. In essence, the MLPF algorithm combines information from tracks and calorimeter clusters to reconstruct charged and neutral hadron, electron, photon and muon candidates.
Further developments of AI models in CoE RAISE have the potential to greatly impact the field of High Energy Physics by efficiently processing the very large amounts of data that will be produced by particle detectors in the coming decades.
In addition, the large scale, parallelizable and resource efficient hyperparameter search algorithms are model agnostic in their nature and could be widely applicable in other sciences making use of AI, for instance in the use-cases of seismic imaging, remote sensing, defect-free additive manufacturing and sound engineering that are part of CoE RAISE WP4.
During the LHC Run 3 the ALICE online computing farm will process up to 50 times more Pb-Pb events per second than in Run 2. The implied computing resource scaling requires a shift in the approach that comprises the extensive usage of Graphics Processing Units (GPU) for the processing. We will give an overview of the state of the art for the data reconstruction on GPUs in ALICE, with additional focus on the Inner Tracking System detector. A detailed teardown of adopted techniques, implemented algorithms and approaches and performance report will be shown. Additionally, we will show how we support different GPUs brands (Nvidia and AMD) with a single codebase using an automatic code translation and generation for different target architectures. Strengths and possible weaknesses of this approach will be discussed. Finally, an overview of the next steps towards an even more comprehensive usage of GPUs in ALICE software will be illustrated.
RooFit is a toolkit for statistical modelling and fitting, and together with RooStats it is used for measurements and statistical tests by most experiments in particle physics, particularly the LHC experiments. As the LHC program progresses, physics analysis becomes more computationally demanding. Therefore, the focus of RooFit developments in recent years was performance optimization. Recently, much of RooFit's core functionality has been re-implemented to either use GPUs or the vector instructions on the CPU, depending on the available hardware.
This presentation will explain which parts of RooFit are implemented to benefit from these hardware accelerators and demonstrate the performance improvements for typical binned and unbinned fits. An overview of the underlying computation library will be given, illustrating how one can reuse the same code for both GPU and CPU libraries and showing the necessary steps to implement custom pdfs.
We will also talk about which remaining RooFit functionality will be ported to hardware accelerators in the future, e.g. the analytic integration of probability densities. Finally, we will highlight other new RooFit features available in the upcoming release, including new functionality specific to PyROOT and new ways to pass the results of auxiliary measurements to the model.
The analysis of high-frequency financial trading data faces similar problems as High Energy Physics (HEP) analysis. The data is noisy, irregular in shape, and large in size. Recent research on the intra-day behaviour of financial markets shows a lack of tools specialized for finance data, and describes this problem as a computational burden. In contrary to HEP data, finance data consists of time series. Each time series spans multiple hours from the start to the end of a trading session, and is related to others (i.e., multiple financial products are traded in parallel at an exchange).
This presentation shows how ROOT can be used in high-frequency finance analysis, which extensions are required to process time series data, and what the advantages are with regard to high-frequency finance data. We provide implementations for data synchronisation (i.e., zipping multiple files together), iterating over the data sequentially with a mutable state (i.e., each entry updates the state of a financial product), generating snapshots (i.e., resampling data based on the timestamps of the entries), and visualisation. These transformations make it possible to fold time series data into high-dimensional data points, where each data point contains an aggregation of recent time steps. This new dataset removes the need to process data serially as a time series, and instead allows the use of parallelised tools in ROOT, like RDataFrame.
We present two applications of declarative interfaces for HEP data analysis allowing users to avoid writing event loops that simplify code and enable performance improvements to be decoupled from analysis development. One example is FuncADL, an analysis description language inspired by functional programming developed using Python as a host language. In addition to providing a declarative, functional interface for transforming data, FuncADL borrows design concepts from the field of database query languages to isolate the interface from the underlying physical and logical schemas. This way, the same query can easily be adapted to select data from very different sources and formats and with different execution mechanisms. FuncADL is part of a suite of analysis software tools being developed by IRIS-HEP for highly scalable physics analysis for the LHC and HL-LHC. FuncADL is demonstrated by implementing the example analysis tasks designed by HSF and IRIS-HEP. The other example is the domain-specific analysis description language ADL that expresses the physics content of an analysis in a standard and unambiguous way, independent of computing frameworks. In ADL, analyses are described in human-readable plain text files, clearly separating object, variable and event selection definitions in blocks having a keyword-expression structure with a self-descriptive syntax that includes mathematical and logical operations, comparison and optimisation operators, reducers, four-vector algebra and commonly used HEP functions. Two infrastructures are available to render ADL executable: CutLang, a direct runtime interpreter written in C++, using lex and yacc for ADL parsing; and adl2tnm, a transpiler converting ADL into C++ or python code. ADL/CutLang are already used in several physics studies and educational projects, and are adapted for use with LHC Open Data. A growing ADL analysis database is available on GitHub.
This track focuses on computing techniques and algorithms used in the theoretical side of physics research.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
Phenomenological studies of high-multiplicity scattering processes at collider experiments present a substantial theoretical challenge and are increasingly important ingredients in experimental measurements. We investigate the use of neural networks to approximate matrix elements for these processes, studying the case of loop-induced diphoton production through gluon fusion. We train neural network models on one-loop amplitudes from the NJet library and interface them with the Sherpa Monte Carlo event generator to provide the matrix element within a realistic hadronic collider simulation. Computing some standard observables with the models and comparing to conventional techniques, we find excellent agreement in the distributions and a reduced simulation time by a factor of 30.
In this talk, I present the computation of the two-loop helicity amplitudes for Higgs boson production in association with a bottom quark pair. I give an overview of the method and describe how computational bottlenecks can be overcome by using finite field reconstruction to obtain analytic expressions from numerical evaluations. I also show how the method of differential equations allows us to express the answers using a basis of special functions whose numerical values can be readily evaluated at any point in phase space. Finally, I discuss the obstacles of loop computations and potential advances in the field.
This talk is based on arXiv:2107.14733
We propose a novel method for the elimination of negative Monte Carlo event
weights. The method is process-agnostic, independent of any analysis, and preserves all physical observables. We demonstrate the overall performance and systematic improvement with increasing event sample size, based on predictions for the production of a W boson with two jets calculated at next-to-leading order perturbation theory.
We present results for Higgs boson pair production in gluon fusion including both, NLO (2-loop) QCD corrections with full top quark mass dependence as well as anomalous couplings related to operators describing effects of physics beyond the Standard Model.
The latter can be realized in non-linear (HEFT) or linear (SMEFT) Effective Field Theory frameworks.
We show results for both and discuss the impact of different approximations within the SMEFT description.
In view of the null results (so far) in the numerous channel-by-channel searches for new particles at the LHC, it becomes increasingly relevant to change perspective and attempt a more global approach to finding out where BSM physics may hide. To this end, we developed a novel statistical learning algorithm that is capable of identifying potential dispersed signals in the slew of published LHC analyses. The task of the algorithm is to build candidate "proto-models" from small excesses in the data, while at the same time remaining consistent with all other constraints. At present, this is based on the concept of simplified models, exploiting the SModelS software framework and its large database of simplified-model results from ATLAS and CMS searches for new physics.
In this talk, we explain the concept as well as technical details of the statistical learning procedure. A crucial aspect is the ability to construct reliable likelihoods in proto-model space and a robust recipe for how to combine them. We will also discuss various aspects of the test statistic employed in our approach. With the current setup, the best-performing proto-model consists of a top partner, a light-flavor quark partner, and a lightest neutral new particle with masses of about 1.2 TeV, 700 GeV and 160 GeV, respectively, and SUSY-like cross sections; for the SM hypothesis we find a global p-value of 0.19.
FeynCalc is esteemed by many particle theorists as a very
useful tool for tackling symbolic Feynman diagram calculations
with a great amount of transparency and flexibility.
While the program enjoys an excellent reputation
when it comes to tree level and 1-loop calculations,
the usefulness of FeynCalc in multi-loop projects is
often doubted by the practitioners.
In this talk I will report on the upcoming version
of the package aiming to address
these shortcomings. In particular, FeynCalc 10
will introduce a number of new routines that facilitate
two very important steps of almost every multi-loop calculation.
The first one concerns the identification of the occurring
multi-loop topologies including the minimization of their
number by finding suitable mappings between integral families.
The second one deals with the handling (visualization,
expansions, analytic evaluation) of master integrals
obtained after a successful IBP reduction of multiple
integral families.
In FeynCalc 10 these nontrivial operations are implemented
in the form of versatile and easy-to-use functions such as
FCLoopFindTopologyMappings, FCLoopIntegralToGraph or
FCFeynmanParametrize etc. that will be introduced in my
presentation.
The technology of quantum computers and related systems is advancing rapidly, and powerful programmable quantum processors are already being made available by various companies. Long before we reach the promised land of fully fault tolerant large scale quantum computers, it is possible that unambiguous “quantum advantage” will be demonstrated for certain kinds of problems, including problems of interest to physicists. At the same time, advances in the technology of quantum processors are likely to go together with advances in quantum sensing and quantum communications, enabling new kinds of physics experiments.
I provide a perspective on the development of quantum computing for data science, including a dive into state-of-the-art for both hardware and algorithms and the potential for quantum machine learning.
With the upcoming High Luminosity LHC coming online in the near future, event generators will need to generate a similar number of events. Currently, the current estimated cost to generate these events exceeds the computing budget of the LHC experiments. To address these issues, the event generators need to improve their speed. Many different approaches are being taken to achieve this goal. I will cover the ongoing work on implementing event generators on the GPUs, machine learning the phase space, improving unweighting efficiencies, and minimizing the number of negative weight events.
This track includes topics that impact how we do physics analysis and research that are related to the enabling technology.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
Query languages for High Energy Physics (HEP) are an ever present topic within the field. A query language that can efficiently represent the nested data structures that encode the statistical and physical meaning of HEP data will help analysts by ensuring their code is more clear and pertinent. As the result of a multi-year effort to develop an in-memory columnar representation of high energy physics data, the numpy, awkward arrays, and uproot python packages present a mature and efficient interface to HEP data. Atop that base, the coffea package adds functionality to launch queries at scale, manage and apply experiment-specific transformations to data, and present a rich object-oriented columnar data representation to the analyst. Recently, a set of Analysis Description Language (ADL) benchmarks has been established to compare HEP queries in multiple languages and frameworks. In this paper we present these benchmark queries implemented within the coffea framework and discuss their readability and performance characteristics. We find that the columnar queries perform as well or better than the implementations given in previous studies.
Awkward Array 0.x was written entirely in Python, and Awkward Array 1.x was a fresh rewrite with a C++ core and a Python interface. Ironically, the Awkward Array 2.x project is translating most of that core back into Python (leaving the interface untouched). This is because we discovered surprising and subtle issues in Python-C++ integration that can be avoided with a more minimal coupling: we can still put performance-critical code in C++, but also benefit by minimizing the interface between the two languages.
This talk is intended to share what we learned from our experiences: design choices that look innocent but can cause issues several steps later, often only in the context of real applications. The points to be presented are (1) memory management: although Python references can be glued to std::shared_ptr
, cycles through C++ are invisible to Python's garbage collector and can arise in subtle ways, (2) C++ standard library types are not a portable runtime interface, owing to ABI differences, and (3) tracers, at the heart of Python libraries like Dask and JAX, can only be fully leveraged if black-box calls out of Python use basic, universally recognized types: flat arrays, not objects.
The goal of this talk is to call out these issues so that other projects mixing Python and C++ can avoid them in the design stage.
For the last 7 years Accelogic pioneered and perfected a radically new theory of numerical computing codenamed "Compressive Computing", which has an extremely profound impact on real-world computer science [1]. At the core of this new theory is the discovery of one of its fundamental theorems which states that, under very general conditions, the vast majority (typically between 70% and 80%) of the bits used in modern large-scale numerical computations are absolutely irrelevant for the accuracy of the end result. This theory of Compressive Computing provides mechanisms able to identify (with high intelligence and surgical accuracy) the number of bits (i.e., the precision) that can be used to represent numbers without affecting the substance of the end results, as they are computed and vary in real time. The bottom line outcome would be to provide a state-of-the-art compression algorithm that surpasses those currently available in the ROOT framework, with the purpose of enabling substantial economic and operational gains (including speedup) for High Energy and Nuclear Physics data storage/analysis. In our initial studies, a factor of nearly x4 (3.9) compression was achieved with RHIC/STAR data where ROOT compression managed only x1.4.
As a collaboration of experimental scientists, private industry, and the ROOT Team, our aim is to capitalize on the substantial success delivered by the initial effort and produce a robust technology properly packaged as an open-source tool that could be used by virtually every experiment around the world as means for improving data management and accessibility.
In this contribution, we will present our efforts integrating our concepts of "functionally lossless compression" within the ROOT framework implementation, with the purpose of producing a basic solution readily integrated into HENP applications. We will also present our progress applying this compression through realistic examples of analysis from both the STAR and CMS experiments.
Recently, graph neural networks (GNNs) have been successfully used for a variety of reconstruction problems in HEP. In this work, we develop and evaluate an end-to-end C++ implementation for inferencing a charged particle tracking pipeline based on GNNs. The pipeline steps include data encoding, graph building, edge filtering, GNN and track labeling and it runs on both GPUs and CPUs. The ONNX Runtime C++ API is used to run PyTorch deep learning models converted to ONNX. The implementation features an improved GPU-based fixed radius nearest neighbor search for identifying edges and a weakly connected component algorithm for the labeling step. In addition, complete conversion to C++ allows integration with existing tracking software, including ACTS. We report the memory usage, average event latency, and the efficiency and purity tracking performance of our implementation applied to the TrackML benchmark dataset. The GPU-based implementation provides considerable speed-ups over the CPU-based execution and can be extended to run on multiple GPUs.
Neutrinos are particles that interact rarely, so identifying them requires large detectors which produce lots of data. Processing this data with the computing power available is becoming more difficult as the detectors increase in size to reach their physics goals. Liquid argon time projection chamber (LArTPC) neutrino experiments are expected to grow in the next decade to have 100 times more wires than currently operating experiments, and modernization of LArTPC reconstruction code, including parallelization both at data- and instruction-level, will help to mitigate this challenge.
The LArTPC hit finding algorithm, which reconstructs signals from the detector wires, is used across multiple experiments through a common software framework. In this talk we discuss a parallel implementation of this algorithm. Using a standalone setup we find speed up factors of two times from vectorization and 30-100 times from multi-threading on Intel architectures, with close to ideal scaling at low thread counts. This new version has been incorporated back into the framework so that it can be used by experiments. On a serial execution, the integrated version is about 10 times faster than the previous one and, once parallelization is enabled, further speedups comparable to the standalone program are achieved.
To fully take advantage of all levels of parallelism in a production environment, data processing must be done at a high performance computing center (HPC). A HPC workflow is being developed to be used as part of a central processing campaign for LArTPC experiments with the goal to efficiently utilize the available parallel resources within and across nodes, as well as AI algorithms. Further opportunities for algorithm parallelism in the reconstruction and GPU code portability are also being explored.
There are as many different algorithms and methods as their are physicists. Obviously, we can't list them all here, but here are some broad outlines of techniques that fit into this category. Of course, new and novel categories are part of what this conference is looking for.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
The Circular Electron Positron Collider (CEPC) [1] is one of future experiments which aims to study the properties of Higgs boson and to perform searches for new physics beyond the Standard Model. The drift chamber is a design option for the outer tracking detector. With the development of new technology in electronics, employment of primary ionization counting method [2-3] to identify charged particles becomes feasible. In order to evaluate particle identification performance, development of a powerful simulation tool is a necessity, which can be used to precisely simulate the response of the gaseous particle detector.
Combining Geant4 and Garfield++ to do this precise simulation has already been studied[4], in which Geant4 is responsible for the primary particle generation and Garfield++ deals with the drift of ions and electrons, amplification via electron avalanches and final signal generation.This contribution will present the integration of Geant4 with Garfield++ in the Gaudi-based CEPCSW software framework. In this implementation, the model used by Geant4 simulation can be configured as either Geant4 PAI or Heed PAI.
Being extremely time-consuming, it is not practical to add Garfield++ into the CEPC simulation chain. To overcome the barrier, data from the Garfield++ simulation is used to train a fully-connected neural network. And the achieved neural network model is then used to transform the information of an ionized electron to its corresponding hit pulse on the signal wire. After the conversion completes, the waveform is created just by piling up the pulses according to their arrival time. This contribution will introduce the details of the above development. Performance studies show that consistent physics results have been obtained and about 200 times speedup makes it possible to apply the machine-learning method to the full detector simulation.
Reference:
[1] CEPC Study Group Collaboration, M. Dong et al., CEPC Conceptual Design Report: Volume 2 - Physics & Detector, arXiv:1811.10545
[2] Jean-Fran¸cois Caron, et al., Improved Particle Identification Using Cluster Counting in a Full-Length Drift Chamber Prototype, 10.1016/j.nima.2013.09.028
[3] F. Cuna, et al., Simulation of particle identification with the cluster counting technique. arXiv:2105.07064
[4] Dorothea Pfeiffer, et al. Interfacing Geant4, Garfield++ and Degrad for the Simulation of Gaseous Detectors. 10.1016/j.nima.2019.04.110
The AI for Experimental Controls project at Jefferson Lab is developing an AI system to control and calibrate a large drift chamber system in near-real time. The AI system will monitor environmental variables and beam conditions to recommend new high voltage settings that maintain consistent dE/dx gain and optimal resolution throughout the experiment. At present, calibrations are performed after data has been taken and require a considerable amount of time and attention from experts. The calibrations require multiple iterations and depend on accurate tracking information. This work would reduce the amount of time and or data that needs to be calibrated in an offline setting. Our approach uses environmental data, such as atmospheric pressure and gas temperature, and beam conditions, such as the flux of incident particles as inputs to a Sequential Neural Network. For the data taken during the latest run period, the SNN is successfully able to predict the existing gain correction factors to within 4%. This talk will briefly describe the development, testing, and future plans for this system at Jefferson Lab.
Jiangmen Underground Neutrino Observatory (JUNO), located at the southern part of China, will be the world’s largest liquid scintillator(LS) detector. Equipped with 20 kton LS, 17623 20-inch PMTs and 25600 3-inch PMTs, JUNO will provide a unique apparatus to probe the mysteries of neutrinos, particularly the neutrino mass ordering puzzle. One of the challenges for JUNO is the high precision vertex and energy reconstruction for reactor neutrino events. This talk will cover both traditional event reconstruction algorithms based on likelihood functions and novel algorithms utilizing machine learning techniques in JUNO.
Particle identification is one of most fundamental tools in various particle physics experiments. For the BESIII experiment on the BEPCII, the realization of numerous physical goals heavily relies on advanced particle identification algorithms. In recent years, the emerging of quantum machine learning could potentially arm particle physics experiments with a powerful new toolbox. In this work, targeting at the muon/pion discrimination problem at BESIII, we have developed a quantum SVM classifier under the Noisy Intermediate-Scale Quantum (NISQ) device. By studying and optimizing various encoding circuits, the quantum SVM trained with the BESIII simulation data shows comparable discrimination power than other traditional machine learning models. This has demonstrated the potential of using quantum machine learning techniques to form a new approach for particle identification in particle physics experiments. In this talk, we present the application of the quantum SVM for particle identification at BESIII, and demonstrate how to construct and optimize the quantum kernel. Furthermore, we present results obtained from the noisy simulator as well as the small-scale hardware to show the potential advantage of the quantum SVM algorithm.
The particle-flow (PF) algorithm at CMS combines information across different detector subsystems to reconstruct a global particle-level picture of the event. At a fundamental level, tracks are extrapolated to the calorimeters and the muons system, and combined with energy deposits to reconstruct charged and neutral hadron candidates, as well as electron, photon and muon candidates.
In light of the upcoming Run 3 as well as the future Phase-2 running conditions with increased pileup and a more fine-grained detector, it is necessary to revisit both the physics and computational performance of the PF algorithm.
We study a machine-learned approach for PF (MLPF) reconstruction in CMS with possible application in offline reconstruction at CMS towards the end of Run 3. The tracks and calorimeter clusters in the event are processed by a graph neural network to reconstruct the full list of PF candidates. Training is carried out on simulated samples created using the CMS software framework (CMSSW) and the full detector simulation model.
We report the technical details of the simulation and training setup, as well as initial physics and computational performance characteristics when the MLPF model is used for inference in CMSSW and interfaced with the downstream reconstruction of high-level objects such as jets and missing transverse momentum.
We present an application of anomaly detection techniques based on deep recurrent autoencoders to the problem of detecting gravitational wave signals in laser interferometers. Trained on noise data, this class of algorithms could detect signals using an unsupervised strategy, i.e., without targeting a specific kind of source. We develop a custom architecture to analyze the data from two interferometers. We compare the obtained performance to that obtained with other autoencoder architectures and with a convolutional classifier. The unsupervised nature of the proposed strategy comes with a cost in terms of accuracy, when compared to more traditional supervised techniques. On the other hand, there is a qualitative gain in generalizing the experimental sensitivity beyond the ensemble of pre-computed signal templates. The recurrent autoencoder outperforms other autoencoders based on different architectures. The class of recurrent autoencoders presented in this paper could complement the search strategy employed for gravitational wave detection and extend the reach of the ongoing detection campaigns.
This track focuses on computing techniques and algorithms used in the theoretical side of physics research.
More information of the scientific programme: https://indico.cern.ch/event/855454/program
We present the mixed QCD-EW two-loop virtual amplitudes for the neutral current Drell-Yan production. The evaluation of the two-loop amplitudes is one of the bottlenecks for the complete calculation of the NNLO mixed QCD-EW corrections. We present the computational details, especially the evaluation of all the relevant two-loop Feynman integrals using analytical and semi-analytical methods. We perform the subtraction of universal infrared singularities and present the numerical evaluation of the hard function.
We present an application of major new features of the program pySecDec, which is a program to calculate parametric integrals, in particular multi-loop integrals, numerically.
One important new feature is the ability to integrate weighted sums of integrals in a way which is optimised to reach a given accuracy goal on the sums rather than on the individual integrals, another one is the option to perform asymptotic expansions.
These new assets of the program represent an important step towards the efficient evaluation of multi-loop amplitudes in a largely automated way. The poster will illustrate them in a pedagogical example, through the calculation of the one-loop amplitude for Higgs plus jet production in the gluon channel, mediated by a top quark loop. Numerical results in the heavy top limit, which can also be obtained with the new version of pySecDec, will be shown as well.
A high-precision calculation of lepton magnetic moments requires an evaluation of QED Feynman diagrams up to five independent loops.
These calculations are still important:
1) the 5-loop contributions with lepton loops to the electron g-2 are still not double-checked (and can potentially be sensitive in experiments);
2) there is a discrepancy in different calculations of the 5-loop contribution without lepton loops to the electron g-2;
3) the QED uncertainty in the muon g-2 is estimated as negligibly small relative to the hadronic one, but there are a lot of questions for those estimations.
Using known universal approaches (based on dimensional regularization and so on) for these purposes leads to an enormous amount of computer time required. To make these calculations feasible it is very important to remove all divergences before integration and avoid limit-like regularizations at intermediate stages. The direct usage of BPHZ is not possible due to infrared divergences of complicated structure.
I had already developed a method of divergence elimination some years ago. That method allowed us to double-check the 5-loop QED contribution without lepton loops to the electron g-2.
A new method of doing this will be presented. Both methods are based on applying linear operators to Feynman amplitudes of ultraviolet divergent subdiagrams. This is similar to BPHZ; the difference is in the operators used and in the way of combining them. Both methods are suitable for calculating the universal (mass-independent) QED contributions to the lepton anomalous magnetic moments: they are equivalent to the on-shell renormalization; the final result is obtained by summation of the diagram contributions; no residual renormalization is required.
However, the new method has some advantages relative to the old one:
1) it works for calculating the contributions dependent on the relations of particle masses (for example, muon and electron);
2) it preserves gauge-invariant classes of diagrams with lepton loops (including mass-dependent ones).
The new method is a next step towards the dream of the general case regularization-free perturbative calculation in quantum field theory.
An algorithm for the spinor amplitudes with massive particles is implemented in the SANC computer system framework.
Procedure for simplification of the expressions with spinor products is based on little group technique in six-dimensional space-time.
Amplitudes for bremsstrahlung processes e+e+\to (e+e+/mu+mu-/HZ/Zgamma/gamma gamma) + gamma are obtained in gauge-covariant form analytically eliminating all gauge-fixing parameters. Numerical integration is performed by ReneSANCe Monte-Carlo event generator.
In recent work we computed 4-loop integrals for self-energy diagrams with 11 massive internal lines. Presently we perform numerical integration and regularization for diagrams with 8 to 11 lines, while considering massive and massless cases. For dimensional regularization, a sequence of integrals is computed depending on a parameter ($\varepsilon$) that is incorporated via the space-time dimension, and approaches zero. We consider diagrams where the leading term in the expansion is in $1/\varepsilon^2$ or in $1/\varepsilon$ or finite. The numerical integration methods include non-adaptive, double exponential integration, and Quasi-Monte Carlo approximations with embedded lattice rules implemented in CUDA C for acceleration on GPU, as well as adaptive integration layered over MPI. The leading term coefficient of $1/\varepsilon^2$ or $1/\varepsilon$ in the integral expansion is obtained via linear or nonlinear extrapolation as epsilon tends to zero.
Over the past decades nuclear physics experiment has seen a drastic increase in complexity. With the arrival of second generation radioactive ions beams facilities all over the world, the run for exploring more and more exotic nuclei is raging. The low intensity of RI-beams require more complex setup, covering larger solid angle, and detecting a wider variety of charged and neutral particles. Design, construction and operation of the variety of complex instruments used in such experiments require more and more software development. The short lifetime of experimental setup and the endless recombination of instruments demands a strong methodology. As the community is shifting to this new paradigm, the quest for the optimum framework is becoming central in the field. In this outlook I will introduce the specificity of the nuclear physics community, technical needs of such framework, and give an overview of the existing one, with an emphasize on the difficult balance between computing performances, versatility and integration with other framework.
Artificial Neural Networks in High Energy Physics: introduction and goals
Nowadays High Energy Physics (HEP) analyses take generally advantages of the implementation of Machine Learning techniques to optimize the discrimination between signal and background, preserving as much signal as possible. Running a classical cut-based selection would imply a severe reduction of both signal and background candidates as well, which would turn to be a quite inefficient choice especially when signal events are rare as it usually happens performing beyond Standard Model studies.
HEP Multivariate Analysis (MVA) focuses on using a pre-determined set of independent variables optimally combined to build discriminants which could effectively separate signal from background.
In this context, an Artificial Neural Network naturally implements a MVA since it is a complex computing system which mimics the biological neural networks receiving these variables as input, training and testing on them and, eventually, producing a single output for binary classification problems. Generally, a neural network is composed of nodes organized in layers. Each node receives either a feature of the problem or a weighted sum of the previous layer node output. Each neural network presents different parameters (i.e. number of layers, nodes, activation functions, etc) and hyper-parameters (parameters set manually in order to help the estimate model parameters). These networks usually implement a model which performs an iterative process (the algorithm) whose aim is to minimize a given loss function that represents the distance of the network response from the actual class of the events. Managing big amount of data for these classification problems via Deep Neural Networks (DNNs) can ensure that the network will correctly classify data never seen before.
The production of Double Higgs via Vector Boson Fusion
In our work we will show the implementation and optimization of a DNN for signal and background event classification.
For this study we used Monte Carlo generated events from non-resonant Higgs boson pair production analysis at the energies of the LHC, where one of the Higgs bosons decays into the four-lepton final state and the other one decays into a pair of b quarks. The signal and background datasets were generated to run the analysis on data corresponding to the integrated luminosity reached by the LHC experiment at CERN during the full Run II period using proton-proton collisions at a center of mass energy of 13 TeV. Furthermore, the performance of a Random Forest classifier, usually considered a very versatile and efficient method for these kind of problems, will be studied, compared with the DNN and discussed. The results will be given exploiting different metrics, providing accuracy and purity times efficiency distributions, confusion matrices and ROC curves for both models.
Innovation and further improvements
The discovery of the Higgs boson at the Large Hadron Collider in 2012 opened a new frontier in HEP both for Standard Model (SM) and Beyond Standard Model (BSM) scenarios. Thus, a new era of ambitious high luminosity studies has been opened up. After the precise measurement of the main parameters of the SM Higgs, one of the most important determination to be accomplished is the measurement of the Higgs self-couplings, which are strictly related to the shape of the Higgs potential. The Higgs self-coupling studies clearly involve the investigation processes which have a pair of Higgs bosons. In contrast to what happens for the dominating production mechanism of gluon-gluon fusion, the production of these two Higgs bosons via vector-boson fusion (VBF) turns to be a particularly important process for the determination of the triple-Higgs coupling. In fact, in VBF the Higgs bosons are produced at leading order from heavy gauge bosons which are radiated off two quarks which can be used as tags for jets to simplify the experimental identification and measurement. Thus, the innovation for this study is related both to the technological point of view and to the originality of the physics analysis signature considered. In fact, differently from the single Higgs production modes widely explored and studied in the Run I and II at the LHC, the double Higgs boson production via VBF in the 4lepton+ 2 b-jet final state was not yet investigated. This was mainly due to the small value of its cross section weighted with the branching ratios (for the HH production via VBF, with the Higgs mass set to its best fit value of 125,09 GeV, the cross section at 13 TeV is $\sim1.723 fb$ and the corresponding BRs are $2.79\times10^{-4}$ for $H\rightarrow{ZZ}\rightarrow{4l}$, with $l=e,\mu,\tau$, and $5.75\times10^{-1}$ for $H\rightarrow{bb}$), thus requiring an exclusive event selection in order to efficiently perform a background rejection.
Benefits and Feasibility of the study for teaching purpose
In addition to the scientific originality of the work itself mentioned in the previous section and its major importance for a better understanding of SM Higgs potential, from these studies several exercises can be derived for both Analysis and Computing Schools, ensuring a deeper understanding of both the particle physics itself and the use of the most suitable ML techniques. As an example, we studied the problem of hyper-parameters tuning using cross validation and other approaches, trying to understand which is the most performing one for the different models. These techniques have been already applied on a different analysis by the authors and submitted to Hackathon INFN 2021 Competition, targeting master and Ph.D. students. The importance of practicing for learners during hands-on tutorials using the very latest physics analyses seems to be a quite remarkable matter, especially due to the application of different but also interconnected fields such as Particle Physics, Artificial Intelligence, Big Data Analytics and the optimization of computing resources using High Performance Computing Clusters.
NA61/SHINE is a high-energy physics experiment operating at the SPS accelerator at CERN. The physics programme of the experiment was recently extended, requiring a major upgrade of the detector setup. The main goal of the upgrade is to increase the event flow rate from 80Hz to 1kHz by exchanging the read-out electronics of the NA61/SHINE main tracking detectors (Time-Projection-Chambers - TPCs). As the amount of collected data will increase significantly, a tool for online noise filtering is needed. The standard method is based on the reconstruction of tracks and removal of clusters which do not belong to any particle trajectory. However, this method takes a substantial amount of time and resources.
In this talk, a novel approach based on machine learning methods will be presented. It is a promising solution which can reduce computing time. Two techniques are considered: usage of decision trees and deep neural networks. For the second method, different neural network architectures were compared, including convolutional neural networks. The two approaches were tested using previously collected NA61/SHINE data. The preliminary results of cluster classification by these models show a classification efficiency of about 90%. The performance of both methods will be discussed and compared. Integrated algorithms with NA61/SHINE software will be used in the future data-taking campaigns.
CYGNO is developing a gaseous Time Projection Chamber (TPC) for directional dark matter searches, to be hosted at Laboratori Nazionali del Gran Sasso (LNGS), Italy. CYGNO uses He:CF4 gas mixture at atmospheric pressure and relies on Gas Electron Multipliers (GEMs) stack for the charge amplification. Light is produced by the electrons avalanche thanks to the CF4 scintillation properties and is then optically read out by a high-resolution scientific CMOS Camera (sCMOS) and Photo-Multiplier Tubes (PMT). sCMOS are designed for low readout noise, uniformity, and linearity and are therefore capable to track particles down to O(keV) energies. These high-resolution images (2D event projection) are combined with the PMT signal (relative z coordinate information) to obtain 3D reconstruction, with the aim of particle identification and to determine track direction of arrival.
sCMOS images are very well suited to be analysed with Machine Learning techniques (using Convolutional Neural Networks) because of their high granularity and low noise. We will present the CYGNO features and achieved experimental performance, and then focus on the MonteCarlo sCMOS images simulation, reconstruction algorithms to identify and track particles, and the use of Convolutional Neural Networks to classify them into different energy classes of Electron and Nuclear recoils.
In the absence of new physics signals and in the presence of a plethora of new physics scenarios that could hide in the copiously produced LHC collision events, unbiased event reconstruction and classification methods have become a major research focus of the high-energy physics community. Unsupervised machine learning methods, often used as anomaly-detection methods, are trained on Standard Model processes and should indicate if a collision event is irreconcilable with the kinematic features of Standard Model events. I will briefly review popular unsupervised neural network methods proposed for the analysis of high-energy physics collision events. Further, I will discuss how physics principles can guide such methods and how their susceptibility to systematic uncertainties can be curbed.
Various meetings for local particpants: wrap ups, discussion, and business purposes