Fast Machine Learning for Science Conference 2026

Session

Contributed Talks

31 Aug 2026, 16:00

100. A Comparative Study of Neural-Network Architectures for Embedded On-Chip Inference For Future MAPS Pixel Detectors

Gian Michele Innocenti (Massachusetts Inst. of Technology (US))

Presentation

On-detector data reduction has become a primary constraint for next-generation monolithic active pixel sensors (MAPS) operated in continuous readout, which are used in particle physics to realize high-resolution, low-material-budget tracking and vertexing detectors. As pixel pitches decrease and hit rates increase, the bandwidth, power, and material-budget cost of this data transport is...

9. A-Global Holoscan-based Scouting Trigger

Noah Clarke Hall (CERN), Ramya Gurunathan, Vila Andela Petrovic (UCL)

Presentation

High Energy Physics experiments at the energy and intensity frontier face O(Tbps) data rates that must be processed with microsecond latency. To handle the large amount of data, a two-stage selection strategy is typically deployed to select and record the most interesting events in real time, where the first stage is based on custom electronics and the second one on a heterogeneous CPU/GPU...

61. Accelerating the Full Particle Transformer on the AMD Versal AI Engine

Yuheon Joh (University of California San Diego)

Presentation

While transformer-based models such as the Particle Transformer have achieved strong performance offline, deploying the full model in a low-latency, resource-constrained trigger system remains an open problem. We extend the previous implementation of a quantized, integer-only transformer for jet tagging on the AMD Versal AI Engine by adding support for LayerNorm and GELU and implementing the...

21. Alkaid: A Compiler Framework for Low-Latency Kernels on Hardware

Chang Sun (California Institute of Technology (US))

Presentation

Low-latency machine-learning and data-processing pipelines often contain static dataflow kernels that require fine-grained bitwidth control, arithmetic optimization, and fast hardware performance estimation. This works introduce Alkaid, a free and open-source domain-specific compiler that translates sub-microsecond latency dataflow kernels into platform-agnostic Register-Transfer Level (RTL)...

84. An Asteroseismic Inversion Framework Using Physics-Informed Neural Networks to Estimate Mixing Length Parameters in Evolved Stellar Grids

Pritam Singh (Wichita State University)

Presentation

Determining the internal physical structures of evolved stars remains a central challenge in stellar astrophysics. Standard procedures involve combining 1D stellar evolution tracks from codes like MESA with linear perturbation data from GYRE to reproduce observed stellar oscillation spectra. Although these numerical techniques are highly accurate, mapping observable surface frequencies back to...

97. An extensible transformer architecture implementation on AMD AI-engines for ultra-low-latency stream-data processing

Elias Leutgeb (CERN), Fabian Helmberger (Vienna University of Technology (AT)), Thomas Owen James (CERN)

Presentation

Deep learning techniques are being used for event classification, reconstruction, and triggering in the first stages of data selection and processing at collider experiments, such as CMS and ATLAS at the LHC. Thus far, the networks being used for low-latency inference are predominately based on fully-connected architectures. On the other hand, transformer-based ML models have demonstrated...

66. Autoregressive Models with HGQ and hls4ml on Altera FPGAs

Lauri Antti Olavi Laatu (Imperial College (GB))

Presentation

Autoregressive generative models form the basis of modern LLMs. These models utilize a transformer architecture which requires caching the previous attention matrix (KV-Cache) for efficient inference. FPGAs are uniquely capable of high throughput due to their on-chip memory for both weights and biases as well as for KV-Cache. We present a GPT based autoregressive model trained using High...

76. Capturing Neutrinos and New Physics at ProtoDUNE with a Machine Learning-Based Trigger Algorithm

Ciaran Hasnip (CERN)

Presentation

The Deep Underground Neutrino Experiment (DUNE) will deploy massive liquid argon time projection chamber (LArTPC) detectors to measure the properties of neutrinos with unprecedented precision. Currently, two full-scale prototypes of the DUNE far detectors (FDs) are at CERN, which are called the ProtoDUNE detectors. Recent studies indicate that these prototypes could have the potential to...

88. Collide-2V Challenge: Edge AI and Foundation Model Development for High-Energy Physics

Benedikt Maier (Imperial College (GB))

Presentation

Machine learning is increasingly central to both offline data analysis and applications on edge systems at the Large
Hadron Collider (LHC). However, progress is hampered by a lack of standardised benchmarks that simultaneously address
the diverse requirements of high-energy physics, from high-precision offline inference to ultra-low-latency inference on
FPGA hardware. This paper introduces...

81. Design Rules for Extreme-Edge Scientific Computing on AI Engines

Zhenghua Ma (University of California San Diego)

Presentation

Extreme-edge scientific applications use machine learning models to analyze sensor data and make real-time decisions.
Their stringent latency and throughput requirements demand small batch sizes and require that model weights remain fully on-chip.
Spatial dataflow implementations are common for extreme-edge applications. Spatial dataflow works well for small networks, but it fails to scale...

98. Distilling Normalizing Flows for Real-Time Anomaly Detection at the LHC

Noah Clarke Hall (CERN), Tara Tahseen

Presentation

Normalizing flows are principled anomaly detectors, selecting anomalies using a probabilistic per-event likelihood. Extreme latency and resource constraints have prevented the deployment of flow likelihoods within the hardware triggers at the Large Hadron Collider. We bypass these limitations by distilling the likelihood from a large normalizing flow into lightweight student estimators...

68. Distributed Machine Learning for low-latency triggering and reconstruction in future collider experiments

Paul Devouge (Université Paris-Saclay (FR))

Presentation

In modern collider experiments, rare and interesting physics processes must be identified in real time while coping with enormous data rates from a large number of detector sensors. Reconstruction algorithms and trigger decisions already need to be made within microseconds, reducing the data volume by orders of magnitude with very low latency. At next generation particle accelerators such as...

3. Efficient Attention Architectures for Real-Time Jet Tagging at the LHC: SAL-T and PHAT-JeT

Alan Xia

Presentation

Transformer-based jet taggers achieve state-of-the-art classification accuracy at the LHC but remain too computationally expensive for the FPGA-based Level-1 trigger systems planned for the HL-LHC upgrade. We present two complementary attention architectures, SAL-T and PHAT-JeT, that close this gap by introducing physics-motivated structural priors into the attention mechanism while preserving...

35. Evolution of Real-Time Anomaly Detection in the CMS Trigger System: From Run 3 to Phase-2

Sabrina Giorgetti (CERN)

Presentation

During Run 3 of the LHC, the CMS experiment introduced real-time anomaly detection algorithms for online event selection for the first time. The deployment of AXOL1TL and CICADA demonstrated the feasibility of signal-agnostic autoencoder models for ultra-low-latency inference in Level-1 trigger hardware. Building on AXOL1TL, a second anomaly-detection layer has been explored for the...

62. Fast Accelerator Control from Learned Surrogates

Andrew Whitbeck (Fermi National Accelerator Lab. (US))

Presentation

We present a surrogate-based learning framework for optimizing control policies in two particle accelerator applications: Lorentz Force Detuning (LFD) compensation in superconducting radio-frequency (SRF) cavities, and spill intensity regulation for the Mu2e experiment at Fermilab. In both systems, physical dynamics are complex and nonlinear, motivating data-efficient surrogate models that...

64. GASP: Cheap Uncertainty Quantification with Logic Neural Networks

Liv Helen Vage (Princeton University (US))

Presentation

Machine learning models are increasingly deployed where knowing when to trust them matters, yet uncertainty quantification (UQ) is often avoided since it can be costly to run. We show that this cost reduces dramatically for fast architectures which reduce to binary computations, such as logic neural networks. Replacing floating-point arithmetic with boolean gates makes inference fast and...

11. GELATO: A Generic Event-Level Anomalous Trigger Option for ATLAS

Adrian Chitan (IFIN-HH (RO))

Presentation

The absence of beyond-Standard-Model physics discoveries at the LHC suggests that new physics may evade conventional trigger strategies. Existing ATLAS triggers control data collection rates either by applying energy thresholds or by targeting event topologies specific to particular models. Unsupervised machine learning offers a compelling alternative via anomaly detection, enabling a...

50. HEPTv2: End-to-End Efficient Point Transformer for Charged Particle Reconstruction

Yuan-Tang Chou (University of Washington (US)), Jan-Frederik Schulte (Purdue University (US))

Presentation

Charged-particle tracking is a core reconstruction task in high-energy physics that directly impacts all reconstruction and physics studies. At the High-Luminosity Large Hadron Collider (HL-LHC), this must be done under much higher pile-up while preserving both tracking quality and computational efficiency. While existing graph-based approaches achieve physics performance comparable to...

14. Implementation of GPU Tracking as-a-Service for the ATLAS Event-Filter

Adrian Chitan (IFIN-HH (RO))

Presentation

Particle tracking at Large Hadron Collider (LHC) experiments is a crucial component of particle reconstruction, yet it remains one of the most computationally challenging tasks in this process. As we approach the High-Luminosity LHC (HL-LHC) era, the complexity of tracking is expected to increase significantly. Leveraging coprocessors such as GPUs presents a promising solution to the rising...

27. IRONSmith: A Visual Dataflow Design Environment for AMD Ryzen AI NPUs

Brock Sorenson (Arizona State University)

Presentation

IRONSmith: A Visual Dataflow Design Environment for AMD Ryzen AI NPUs

Brock Sorenson, Samer Ali, Curt Bansil, Aman Arora — Arizona State University

---

Machine learning inference increasingly relies on specialized hardware
accelerators for throughput and power efficiency. Neural Processing Units
(NPUs), such as the AMD Ryzen AI NPU, offer significant ML advantages over
CPUs...

90. Just Furnish Context: Autonomous Agents Compress the HEP Analysis Cycle from Years to Hours

Eric Anton Moreno (Massachusetts Institute of Technology (US))

Presentation

Experimental high energy physics analysis is traditionally a multi-year effort dominated by repetitive code development that demands little physics insight. We show that LLM-based AI agents can already autonomously execute substantial portions of this pipeline, collapsing the implementation bottleneck to roughly ten hours of wall-clock time. We present Just Furnish Context (JFC), a...

57. KrIO: Standardized Hardware for Accelerated Science

Noah Paladino (Massachusetts Inst. of Technology (US))

Presentation

As low-latency machine learning techniques are adopted in data collection and processing across various domains, a standard hardware platform becomes necessary to ease adoption. While many FPGA manufacturers provide a few development kits to demonstrate device capabilities, most of these lack the raw inputs necessary to cost-effectively build custom data acquisition systems. We present KrIO, a...

65. MACHINE LEARNING MODELS FOR ICE FORECASTING IN THE GREAT LAKES

Dr Jia Wang (NOAA Great Lakes Environmental Research Laboratory)

Presentation

Accurate Great Lakes ice forecasting is essential for maritime operations, coastal protection, and offshore engineering, yet traditional physics-based models like FVCOM–CICE remain computationally expensive. This study develops a deep neural network surrogate trained on an extensive FVCOM–CICE hindcast using ERA5 2-m air temperature and FVCOM surface water temperature as physically motivated...

63. Maskformer models for particle reconstruction in high-granularity calorimeters

Matteo Marchegiani (Carnegie-Mellon University (US))

Presentation

The calorimeters employed in large experiments at the Phase-1 LHC were designed to cope with the radiation levels and hit multiplicities expected in proton-proton collisions with an average of 20–50 interactions per bunch crossing. However, the High-Luminosity LHC (HL-LHC) will operate with pileup levels of up to 200 interactions per crossing. The radiation environment is especially...

2. Memristive tabular variational autoencoder for compression of analog data in high energy physics

Tae Min Hong (University of Pittsburgh (US))

Presentation

We present an implementation of edge AI to compress data on an in-memory analog content-addressable memory (ACAM) device. A variational autoencoder is trained on a simulated sample of energy measurements from incident high-energy electrons on a generic three-layer scintillator-based calorimeter. The encoding part is distilled into tabular format by regressing the latent space variables using...

31. On-chip probabilistic inference for charged-particle tracking at the sensor edge

David Jiang (Univ. Illinois at Urbana Champaign (US))

Presentation

Modern scientific instruments operate under increasingly extreme constraints on bandwidth, latency, and power. Inference at the sensor edge determines experimental data collection efficiency by deciding which information to save for further analysis. Particle tracking detectors at the Large Hadron Collider exemplify this challenge: pixelated silicon sensors generate rich spatiotemporal...

8. Passage of particles through matter and the effective straggling-function: High-fidelity accelerated simulation via Physics-Informed Machine Learning

Rotem Dover (Department of Particle Physics and Astrophysics, Weizmann Institute of Science)

Presentation

High-fidelity simulation of particle–matter interactions provides the essential theoretical reference for diverse physics disciplines, yet generating synthetic datasets at the scale of current and future experiments has become prohibitive. Here, we introduce PHIN-GAN, a novel physics-informed generative adversarial network designed to address this challenge. We derive a set of analytical...

46. pyrtlnet: An End-to-End Python Workflow for Quantized Inference Hardware Development

Timothy Sherwood (UC Santa Barbara)

Presentation

As quantized neural network inference moves from cloud to edge, the demand for custom hardware accelerators tailored to specific models, precisions, and deployment constraints continues to grow. While frameworks for training and deploying quantized models in software are mature, the path from a trained model to working custom hardware remains steep and requires expertise spanning machine...

73. Real-Time Axion Dark Matter Signal Detection with a WaveNet Architecture Deployed on FPGA

Ms Akbota Assan (University of California, San Diego)

Presentation

The search for axion dark matter requires recovering extremely weak, narrowband signals buried in noise, often under stringent latency and throughput constraints at the data-acquisition stage. We present a WaveNet-based architecture for denoising and detecting axion-like signals on the TIDMAD benchmark, derived from the MIT ABRACADABRA experiment. The model uses dilated causal convolutions and...

19. Real-Time Detection of Charge Jumps in Superconducting Qubits with a Convolutional Neural Network

Jose Daniel Gaytan Villarreal (Carnegie-Mellon University (US))

Presentation

Ionizing radiation from cosmic rays and ambient radioactivity induces charge jumps in superconducting qubits, causing correlated errors that challenge fault-tolerant quantum computing while simultaneously providing a detection signature for quantum sensing applications. Current detection methods operate offline, providing only a post-hoc record of charge events. By the time a jump is...

67. Real-time Event-level Zero-shot Learning at 40 MHz

Leo Rozanov

Presentation

Zero-shot learning (ZSL) refers to the ability of a model to be able to classify unseen labels. We present High Granularity Quantization (HGQ) Linformer based event-level ZSL architecture model trained using contrastive methods on the Collide-2V dataset. The output of the model is an embedding vector which enables classification through ZSL. Different physics processes occupy unique regions in...

22. Real-Time Plasma State Estimation via FPGA-Accelerated Quantized Recurrent Probabilistic Neural Networks

Aiken Xie (Columbia University), Jose Daniel Gaytan Villarreal (Carnegie-Mellon University (US))

Presentation

We evaluate the FPGA deployment of a quantized recurrent probabilistic neural network (RPNN) online state estimator for real-time tokamak control. To navigate the strict resource and latency limitations of fusion environments, we detail a workflow utilizing quantization-aware training (QAT) via the Brevitas framework, followed by translation into an efficient FPGA implementation using a...

32. Real-time Quantum Inspired Algorithms for Anomaly Detection in Collider Triggers

Sagar Addepalli (SLAC National Accelerator Laboratory (US))

Presentation

Real-time Quantum Inspired Algorithms for Anomaly Detection in Collider Triggers
Anomaly detection algorithms deployed at the edge of particle collider experiments offer a model-agnostic approach to search for signs of new physics, complementing typical topology- or energy-driven selections. Tensor networks are a class of quantum-inspired machine learning models that represent information in...

75. Scaling Out hls4ml for ASICs: A Surrogate Model for Neural Network Accelerator Synthesis Metrics

Arghya Ranjan Das (Purdue University)

Presentation

Custom ASIC accelerators offer significant power and performance advantages for machine learning in scientific and edge computing; a driving example is superconducting qubit readout, where moving real-time classification of qubit states from room-temperature FPGAs into the cryostat requires custom ASICs on cryo-compatible technology nodes. However, obtaining accurate area and timing requires...

85. SliceFI: An Optimized, Heuristic-Based Fault Injection Tool for Edge Neural Networks

Andy Meza

Presentation

When deployed in edge applications, neural networks (NNs) undergo numerous changes to ensure they adhere to strict power, performance, and size constraints while simultaneously being robust to faults. In prior work, NN robustness is evaluated using a bit-level ranking based on how sensitive an edge NN is to a fault in a given parameter bit. Unfortunately, the fault injection (FI) campaigns...

39. Smartpixels at a 10 TeV Muon Collider: Machine Learning Algorithms for On-Pixel Rejection of Beam Induced Background

Daniel Abadjiev (University of Chicago (US))

Presentation

At a 10TeV muon collider, a high level of beam induced background (BIB) will overlay signal produced from muon collisions, similar to how pile-up at the HL-LHC will overlay signal from proton-proton collisions. On-detector differentiation of BIB from signal would improve performance of the inner pixel tracker. Following the successful development of a prototype “smartpixel” application...

40. Sparse Streaming Neural Decoders for Real-Time Quantum Error Correction

Mr Eric Jin (University of California, Berkeley)

Presentation

Real-time decoding is the binding constraint for fault-tolerant quantum computing: superconducting qubits emit a syndrome round roughly every microsecond, and a decoder that cannot keep pace incurs an exponentially growing backlog. Neural decoders have reached an accuracy regime that classical methods (Union-Find, MWPM, belief propagation) miss, but few have been demonstrated within the...

16. SpikePID: Neuromorphic Cluster Counting in Future Drift Chambers

Ho-Fung Tsoi (University of Pennsylvania)

Presentation

In collider experiments, particle identification (PID) in drift chambers has traditionally relied on the ionization energy loss ($dE/dx$), whose resolution is fundamentally limited by large Landau fluctuations. Instead, cluster counting ($dN/dx$) measures the number of primary ionization clusters, which is Poisson in nature with smaller statistical fluctuations and can offer higher separation...

72. STRIDE: An Automated Design Space Exploration Workflow for Generating AI Training Datasets from Hierarchical Mathematical Kernels

Angelos Ioannou (Lawrence Berkeley National Laboratory), Giuseppe Di Guglielmo (Fermi National Accelerator Laboratory), John Shalf (Lawrence Berkeley National Laboratory)

Presentation

Developing efficient hardware accelerators for mathematical kernels remains a labor-intensive process, traditionally requiring deep expertise in low-level hardware description languages and extensive manual optimization. While high-level synthesis tools and domain-specific hardware generators have improved productivity, the space of achievable designs for any given kernel is vast, and...

52. SuperSONIC: Inference as a Service on Kubernetes and HPC

Dmitry Kondratyev (Purdue University (US))

Presentation

The increasing ubiquity and complexity of machine learning models in scientific workflows motivate the adoption of inference-as-a-service architectures, which enable efficient utilization of hardware accelerators. The SONIC (Services for Optimized Network Inference on Coprocessors) framework provides a blueprint for implementing inference as a service in high-energy physics (HEP),...

53. Surrogate Models for Real-Time Thermal Runaway Prediction in HTS Magnets

Maira Khan (Fermi National Accelerator Laboratory)

Presentation

High-temperature superconducting (HTS) magnets are a key enabling technology for future muon colliders, fusion energy systems, and other high-field scientific facilities. Ensuring safe operation requires rapid estimation of thermal stability margins and thermal runaway thresholds, yet these quantities are typically obtained through iterative numerical solutions of nonlinear heat-balance...

17. Towards FPGA-Deployed Anomaly Detection for Real-Time Neutrino Data Selection in LArTPCs

Seokju Chung (Columbia University)

Presentation

Raw wire readout data from Liquid Argon Time Projection Chambers (LArTPCs) in future experiments such as the Deep Underground Neutrino Experiment (DUNE) will vastly exceed available storage bandwidth, necessitating intelligent online data selection. We present a compressed autoencoder architecture for model-independent anomaly detection operating directly on LArTPC wire waveforms, benchmarked...

12. Towards the end-to-end optimisation of LHC triggers

Adrian Chitan (IFIN-HH (RO))

Presentation

High-energy physics experiments face extreme data rates, requiring real-time trigger systems to reduce event throughput while preserving sensitivity to rare processes. Trigger systems have traditionally been constructed as modular chains of sequentially optimised algorithms, each optimised for a specific local objective with no guarantee of overall optimality.

We instead formulate trigger...

89. Ultrafast On-chip Online Learning via Spline Locality in Kolmogorov-Arnold Networks

Philip Coleman Harris (Massachusetts Inst. of Technology (US))

Presentation

Ultrafast online learning is essential for high-frequency systems, such as controls for quantum computing and nuclear fusion, where adaptation must occur on sub-microsecond timescales. Meeting these requirements demands low-latency, fixed-precision computation under strict memory constraints, a regime in which conventional networks buil on Multi-Layer Perceptrons (MLPs) are both inefficient...

70. XFMT: LUT-based Transformer for microsecond-scale inference

Arianna Cox (Imperial College (GB))

Presentation

Transformers are promising for real-time intelligent systems, but their arithmetic complexity makes microsecond-scale FPGA inference challenging. This paper presents a LUT-based Transformer framework that combines heterogeneous quantization and LUT-aware training to map compact Transformer models into hardware-efficient lookup-table structures. The proposed method jointly optimizes precision,...