Sixth MODE Workshop on Differentiable Programming for Experiment Design

1 Sept 2026, 09:00 → 7 Sept 2026, 18:45 Europe/Athens

OAC conference center, Kolymbari, Crete, Greece.

Lorenzo Arsini ("Sapienza" University of Rome), Muhammad Awais, Angel Bueno, Tommaso Dorigo (Universita e INFN, Padova (IT)), Luigi Favaro (Universite Catholique de Louvain (UCL) (BE)), Andrea Giammanco (Universite Catholique de Louvain (UCL) (BE)), Christian Glaser (Uppsala University), Hamza Hanif (Simon Fraser University (CA)), Lisa Kusch (TU Eindhoven), Tobias Kortus (RPTU University Kaiserslautern-Landau), Gilles Claude Louppe, Pablo Martinez Ruiz Del Arbol (Universidad de Cantabria and CSIC (ES)), Pietro Vischia (Departamento de Física and ICTEA, Universidad de Oviedo), Gordon Watts (University of Washington (US)), Stéphanie Landrain (Université catholique de Louvain)

This is the sixth installment of a series of workshops where we bring together physicists from particle, astroparticle, and nuclear physics, computer science, and mathematics to develop new methods for experiment design and optimal information extraction from data, powered by differentiable programming.

 

This initiative stems from the activities of the MODE Collaboration. MODE stands for "Machine-learning Optimized Design of Experiments".

 

Location:

The workshop will take place at the OAC (https://www.oac.gr/en/) in Kolymbari, Crete (Greece). Information on the accommodation is available in a dedicated page.

At the same page you will find the procedure for young participants to ask for financial support: we have limited availability to cover part of the travel expenses and/or waiving of the conference fee for some selected young participants. To be considered for funding, you will need to submit an abstract for a talk or poster. The funding will be conditional on the delivery of the talk/poster presentation.

Remote attendance is not foreseen. We want to create the spirit of a scientific retreat, where serendipitous conversations lead to new ideas and collaborations.

Minimal schedule:

• 1 September 2026: arrival day (evening, dinner included in the fee)
• 2--6 September 2026: workshop sessions
• 7 September 2026: departure day (morning, breakfast included in the fee)

Please account for different timezones when consulting the timetable. In particular, Greece is on Eastern European timezone (GMT+2).

Registration and abstract submission:

Please register using the links in the menu to the left.

New registrations close on 20 July 2026.

Registrations must go into “complete” status (i.e. wire transfers received, or confirmation that you will pay by cash at the venue) by 25 July.

 

Overview of the sessions:

• Confirmed keynote speakers

  • TBA

  • TBA

• Lectures and tutorials:

  • Tutorial (TBC)

• Special events:

  • Poster session: prizes will be given to the best posters!

  • Hackathon (TBC): prizes will be given to the winners of the challenge!

• Methods and Tools
• Applications in Muon Tomography
• Applications in particle physics
• Applications in astro-HEP and neutrino physics
• Applications in nuclear physics
• Applications in medical physics and other fields

 

Prizes for special events

There will be a set of prizes for the hackathon, and one for the poster session.

Organising Committee:

You can get in touch with the organising commitee at mode-workshop-organizers@cern.ch.

• Lorenzo Arsini (INFN-Roma and Sapienza Università di Roma)
• Muhammad Awais (INFN-Padova and Luleå Tekniska Universitet)
• Angel Bueno Rodriguez (DLR)
• Tommaso Dorigo (INFN-Padova and Luleå Tekniska Universitet)
• Luigi Favaro (UCLouvain)
• Andrea Giammanco (UCLouvain)
• Christian Glaser (TU Dortmund University)
• Hamza Hanif (Simon Fraser University)
• Lisa Kusch (TU Eindhoven)
• Tobias Kortus (RPTU)
• Gilles Louppe (ULiège)
• Pablo Martinez Ruiz del Árbol (Universidad de Cantabria)
• Pietro Vischia (Universidad de Oviedo)
• Gordon Watts (University of Washington)
• Stéphanie Landrain (secretariat) (UCLouvain)

 

Scientific Advisory Committee:

• Atilim Gunes Baydin (University of Oxford)
• Kyle Cranmer (University of Wisconsin)
• Julien Donini (Université Clermont Auvergne)
• Piero Giubilato (Università di Padova)
• Gian Michele Innocenti (CERN)
• Michael Kagan (SLAC)
• Riccardo Rando (Università di Padova)
• Roberto Ruiz de Austri Bazan (IFIC-CSIC/UV)
• Kazuhiro Terao (SLAC)
• Andrey Ustyuzhanin (SIT, HSE Univ., NUS)
• Christoph Weniger (University of Amsterdam)

 

Funding agencies:

This workshop is partially supported by the joint ECFA-NuPECC-APPEC Activities (JENAA).

This workshop is partially supported by National Science Foundation grant PHY-2323298 (IRIS-HEP).

This workshop is partially supported by the Fund for Scientific Research (F.R.S.–FNRS)

Tuesday 1 September

15:00 Arrival day

Dinner at OAC (included in the fee)

Wednesday 2 September

08:00 Breakfast at OAC (only for people with OAC accommodation)

Registration

Introduction

1 Welcome by the OAC

Speakers: Antonis Kalogerakis, Head of the Institute of Theology & Ecology - Department of the OAC (Orthodox Academy of Crete), Dr Konstantinos Zompas, General Director of the OAC (Orthodox Academy of Crete)

Dr. A. Kalogerakis MODE2026.pdf

2 Welcome and Introduction to the Workshop

Speaker: Prof. Pietro Vischia (Departamento de Física and ICTEA, Universidad de Oviedo)

Sixth MODE Workshop Intro 2026.pdf

Keynote session

Convener: Prof. Pietro Vischia (Departamento de Física and ICTEA, Universidad de Oviedo)

3 AI Agents and the future of HEP Analysis

AI agents are beginning to move beyond writing isolated pieces of code toward executing substantial portions of high-energy physics analyses. Starting from the basic agentic loop - setting goals, using tools, executing code, inspecting results, and iterating - we examine increasingly sophisticated demonstrations, including experiment-software tasks and end-to-end autonomous analyses. This is a story of both progress and limitations: plausible results may contain hidden assumptions, validation gaps, non-determinism, or dependence on rapidly changing models. Scientifically useful agents are likely to require grounded collaboration knowledge, executable tests, independent review, reproducible workflows, and explicit human approval. This leads us to the next "question": how we build the infrastructure and controls needed to make their contributions trustworthy while preserving human responsibility and creativity for scientific progress.

Speaker: Gordon Watts (University of Washington (US))

2026 MODE - AI Agents and the Future of HEP Analysis - 4-3 Attempt.pdf

2026 MODE - AI Agents and the Future of HEP Analysis - 4-3 Attempt.pptx

10:50 Discussion

11:00 Coffee break (included in the fee)

Methods and Tools

4 Toward the codesign of scientific experiments

The design of modern scientific experiments for fundamental physics
entails the choice of geometry, materials, detection techniques, and software reconstruction procedures governed by thousands of parameters. The full optimization of such systems can in most cases be achieved only if all these parameters are considered together in an end-to-end optimization procedure; in particular, hardware and software must be considered together in order to properly account for their interdependence. In this presentation the need for codesign will be discussed using a few examples.

Speaker: Muhammad Awais (University of Padua, Italy & LTU, Sweden)

towards the codesign of scientifc experiments.pdf

5 Beyond (co)design: De novo experiment design and experiment design optimisation in the era of foundation models

Progress in physics has long been driven by ingenious experiments conceived by human experts.
Recently, AI-driven design methods have begun to move beyond tuning a handful of parameters to proposing entirely new experimental layouts. The discovered configurations often challenge established design conventions while matching or even exceeding the performance of human-designed setups.

This talk is based on two papers: in Eur. Phys. J. C 85, 1066 (2025) we frame the concept of intelligence and propose a hybrid intelligence and a framework for the development of hierarchical physics models that may be able to discover on their own new physics laws or concepts. In the other (Nature, https://www.nature.com/articles/s41586-026-10898-6 ), we frame experimental design as a search for optima over a vast space of hardware configurations subject to practical constraints and provide a four-questions framework to make AI-driven design able to work on a spectrum from parameter tuning to de novo discovery, highlighting trade-offs between computational tractability, experimental feasibility, interpretability, and solution reliability.

I then argue that, ultimately, AI-designed experiments might thereby open new ways to explore the universe, and start outlining the strictness requirements that such frameworks must satisfy to be useful and usable.

Speaker: Prof. Pietro Vischia (Departamento de Física and ICTEA, Universidad de Oviedo)

HTML slides (preferred, animations work here)

6 Automatic Differentiation in CMS Combine: From Integration to Physics-Grade Fits

At the previous MODE workshop we presented the beginning of a collaborative effort to enable automatic differentiation (AD) in CMS Combine through RooFit's Clad-based AD framework. In this talk we report on the progress made since: the integration has moved from a proof of feasibility toward a usable capability for realistic CMS statistical models.

We will discuss the RooFit-side developments that closed the gap to Combine's modeling primitives, and the Clad-side compiler work that makes the generated gradients scale: improved activity and to-be-recorded analyses that avoid worst-case tape growth on large likelihoods, and a restructured analysis pipeline informed by an audit against established AD tooling. We will also present the newly-added support for higher-order derivatives (Hessians) in RooFit with Clad.

We will present performance results on cutting-edge Higgs measurements with LHC Run-3 data, comparing AD-based minimization against numerical differentiation in fit time, minimizer robustness, and memory footprint. We will also candidly discuss what did not work: the modeling patterns that resist source transformation, the debugging and validation burden of generated gradients, and what this implies for AD adoption across the wider HEP statistics ecosystem. We aim to turn the lessons from Combine into a reusable recipe for other RooFit-based frameworks.

Speaker: Vassil Vasilev (Princeton University (US))

CladCombine.pdf

13:00 Lunch at OAC (included in the fee)

Free time

15:30 Coffee break included in the fee)

Applications in Particle Physics

7 Co-Designing Hardware and Software in Particle Tracking Detectors

The optimization of charged-particle tracking detectors has traditionally been addressed through sequential workflows, in which the hardware design (sensor geometry, materials, electromagnetic fields, readout schemes) is frozen before the development of reconstruction software (seeding, pattern recognition, track fitting, alignment, identification, and background subtraction) begins. This work formally characterizes when such sequentiality is valid and when it constitutes a suboptimal approximation, employing the framework of block-decomposability and non-factorizable objective functions $f(\mathbf{h},\mathbf{s}) = \Phi(A(\mathbf{h}), B(\mathbf{s}))$ versus genuine couplings of the form $f(\mathbf{h},\mathbf{s}) = A(\mathbf{h}) + B(\mathbf{s}) + \mathbf{h}^{\mathsf{T}} M \mathbf{s}$, where the coupling matrix $M$ encodes the cross-dependence between hardware parameters $\mathbf{h}$.

It is first shown, using a simplified analytical example — a silicon strip tracker with $N$ layers measuring the vertical position of a particle — that in an idealized regime without background noise, inefficiencies, or misalignments, the problem becomes exactly soluble via classical statistical inference (Rao–Cramér–Fréchet bound from Fisher information). In this limit, the optimal hardware (number and positioning of layers) can be determined without any reference to the reconstruction method: the problem is trivially factorizable. However, merely introducing a non-analytic resolution — obtainable only through expensive simulations — and a budget constraint that couples hardware cost to the computational cost of inference suffices to break this separability: co-design then becomes inevitable. This distinction — between problems with closed-form likelihood and problems that depend on simulation-based inference or approximate models — constitutes the conceptual axis of the work.
Building on this framework, six persistent dimensions of hardware–software coupling observed in real trackers are identified and developed:

---

1. MuOnE detector

This silicon strip tracker, designed to measure the differential distribution of muon–electron scattering with a focus on the high-$q^2$ region, illustrates a hardware–software coupling mediated by data-driven extraction of geometric parameters. The layer position $z_i$ and tilt angle $\theta_i$ are not determined by direct metrology, but rather via a global fit over real scattering events, leaving each $z_i$ or $\theta_i$ as a free parameter in successive turns, achieving 1.5 μm precision with $10^7$ reconstructed events. The finding that the dependence of the $q^2$-resolution on the longitudinal vertex position is only relevant under a reconstruction scheme with a constrained vertex (as opposed to an independent fit of the three tracks) motivated a direct hardware redesign: segmenting the beryllium target into carbon-fiber foils of $\le 50~\mu$m spaced in air or vacuum, reducing the $z$-uncertainty and directly impacting the resolution in the high-$q^2$ region of physical interest.

---

2. Sensor positioning and mechanical tolerances

The previous case is generalized: the intrinsic sensor resolution (strips vs. pixels) is modulated by alignment precision and by the axial magnetic field, which determines whether transverse $(x,y)$ or longitudinal $(z)$ resolution dominates. The higher channel density of pixels introduces an overall cost trade-off, while position determination via high-precision laser holographic systems — independently of the data-driven fit — defines a separate Pareto frontier between mechanical/calibration cost and geometric precision.

---

3. Material budget and interaction length $\lambda_I$

Passive material (parameterized in units of radiation length $X_0$ and nuclear interaction length $\lambda_I$) simultaneously couples multiple physical channels: the characteristic multiple-Coulomb-scattering angle

$$ \theta_0 \approx \frac{13.6~\text{MeV}}{\beta p} \sqrt{\frac{x}{X_0}} \left[1 + 0.038 \ln\left(\frac{x}{X_0}\right)\right], $$ the photon conversion probability, and the rate of nuclear interactions that generate fake tracks and spurious secondary vertices. It is argued that the optimal material distribution (few thick layers vs. many thin layers for fixed $x/X_0$) depends explicitly on the track model employed — standard Kalman filter versus treatments with adaptive outlier rejection or smoothing — so that the same hardware configuration yields different effective resolutions and efficiencies depending on the downstream software. It is further discussed how real-time trigger modules ($p_T$-type modules such as those in the CMS Outer Tracker) introduce additional material that degrades offline tracking but enables Level-1 selection based on transverse momentum, rendering the system utility a non-monotonic function of the isolated material budget. --- ### 4. Electron reconstruction and tracker–calorimeter interaction Bremsstrahlung, whose probability grows with $x/X_0$ and with electron energy, generates a strongly non-Gaussian momentum-loss distribution with a long tail of hard emissions. This invalidates the Gaussian process-noise assumption of the standard Kalman filter, motivating the use of Gaussian-Sum Filters (GSF) that represent the electron state as a weighted mixture of momentum hypotheses. It is shown that the marginal benefit of investing in a GSF depends sensitively on whether the material is concentrated in a few well-defined layers or distributed heterogeneously (services, supports), establishing an explicit coupling between hardware architecture and justifiable algorithmic complexity. On the calorimeter side, it is discussed how supercluster construction (e.g., elongated in $\phi$ in CMS) and the longitudinal granularity of the ECAL are co-determined with the expected bremsstrahlung patterns, so that a calorimeter design that is "optimal" under the simplistic assumption of compact showers ceases to be so when realistic reconstruction is considered. --- ### 5. 4D seeding versus pile-up occupancy at the HL-LHC With up to $\mathcal{O}(200)$ interactions per bunch crossing and a vertex time spread $\sigma_{\mathrm{PU}} \sim 180$ ps, precision timing layers (CMS MTD, ATLAS HGTD, $\sigma_t \sim 30$–50 ps) allow hit filtering via a window $\Delta t$ around the predicted time. The scaling of the number of seed combinations is derived as $$ N_{\mathrm{seed}}(\Delta t) \propto \left( \frac{\lambda_0 \, \Delta t}{T_{\mathrm{eff}}} \right)^k $$ for $k$-layer seeds, and the hit acceptance efficiency as $$ p_{\mathrm{hit}}(\Delta t, \sigma_{\mathrm{eff}}) = \operatorname{erf}\left( \frac{\Delta t}{2\sqrt{2}\,\sigma_{\mathrm{eff}}} \right), $$ showing that achieving $\varepsilon_{\mathrm{seed}} \gtrsim 95\%$ for triplet seeds requires $\Delta t \sim (5\text{--}6)\,\sigma_{\mathrm{eff}}$. This establishes that the combinatorial gain depends critically on the ratio $\Delta t / \sigma_{\mathrm{eff}}$ and not on the hardware alone: only if the software dynamically adapts the time window (4D Kalman filter with temporal covariance) does the hardware timing-resolution improvement translate into a superlinear combinatorial reduction, given $N_{\mathrm{seed}} \propto \Delta t^k$. --- ### 6. Converted photons in the $H \to \gamma\gamma$ search The trade-off exploited by CMS in the Higgs discovery is analyzed: the tracker material ($0.4$–$1.0\,X_0$ in the barrel, up to $1.8\,X_0$ in the endcaps) produces photon conversion into $e^+e^-$ pairs with 40–60% probability, enabling high-precision direction reconstruction via track-based conversion fits combined with ECAL clusters. However, the same material increases multiple scattering, bremsstrahlung, and the rate of secondary interactions, degrading the global resolution of unconverted photons. CMS's final decision to prioritize material reduction in Phase-1/Phase-2 upgrades, despite sacrificing conversion statistics, illustrates a case where the quantitative evaluation of the hardware–software coupling directly determined the long-term design strategy.

---

Conclusion

Collectively, these six examples — addressed with a level of formalism ranging from Fisher information to combinatorial scaling arguments — demonstrate that sequential optimization (hardware-first, software-second) is systematically suboptimal in modern trackers, precisely because the relevant inference rarely admits a closed-form likelihood.

Speaker: Hipolito Arturo Riveros Guevara

Mode Presentation [Autoguardado].pdf

8 End-to-end optimization of a Muon Collider Calorimeter

End-to-end approaches to experimental design studies can be developed thanks to the latest advancements in computing capabilities, in particular within Automatic Differentiation. We present here the case study of an Electromagnetic Calorimeter for a proposed future Muon Collider experiment, where muon decays along the beamline provide a significant challenge when attempting to precisely reconstruct events. We illustrate the pipeline development, highlighting the differentiable surrogates set up to simulate both signal and background in the collider environment - in particular the use of Diffusion Models for signal generation. We then proceed to discuss possible utility functions as well as design suggestions, to potentially improve the signal reconstruction efficiency of the experiment.

Speaker: Federico Nardi (Technical University of Munich, Universita' e INFN, Padova (IT))

hist2d_animation.gif

hist_animation.gif

Presentation_MODE_Nardi.key

Presentation_MODE_Nardi.pdf

Presentation_MODE_Nardi.pptx

9 Differentiable Co-Design Simulation Pipeline for Nanophotonic Neuromorphic Processing of Scintillator Detector Light Patterns

Particle detectors are essential in a wide range of applications, from scientific instruments to medical and industrial imaging. Next generation scintillator detectors aim to achieve improved spatio-temporal resolution and particle identification capabilities, which require sub-nanosecond processing of the complex light patterns generated by particles passing through such materials. One promising approach is to perform feature extraction of the light patterns directly near or in the sensor elements using neural networks and neuromorphic architectures whose intrinsic dynamics match the temporal and spatial characteristics of scintillator signals. Relevant features used to infer particle properties are shower centroid, deposited energy, and interaction time.

This work focuses on differentiable co-design modelling and optimization of small clusters of nanophotonic neuron units and nanowire photodetectors with picosecond-scale time constants, based on experimentally motivated device models. The simulation pipeline implements the underlying optoelectronic neuron equations through discrete-time integration in PyTorch, enabling end-to-end differentiable simulation of neuron dynamics, internal memory effects, and optical communication between interconnected nodes. The neuron model incorporates multiple hardware-defined time scales originating from photodetection, charge storage, transistor gating, and light emission processes, allowing the study of how physical device parameters influence temporal information processing. The underlying nanophotonic neuron architecture is based on previously proposed III-V nanowire optoelectronic nodes with internal memory and optical interconnects.

The pipeline includes a configurable data-generation stage in which scintillator-inspired input signals are represented as current pulses in the nanoampere range over nanosecond-scale time windows. Scintillator pulse shapes are modeled as the difference of two exponentials with independently adjustable rise and decay time constants, while pulse ordering, delays, and spatial distribution across input neurons can be varied systematically. This enables the generation of temporally structured datasets for studying basic scintillator light pattern discrimination and feature extraction tasks.

Networks composed of nanophotonic neurons can be instantiated with arbitrary connectivity patterns, including reservoir-style architectures. Because the complete simulation is differentiable, gradients can propagate through both the network topology and the underlying device dynamics, enabling gradient-based optimization of network parameters. Several task-specific objective functions can be defined, including losses designed for classification tasks, based on the integrated optical power emitted by output neurons.

Although inspiration comes from neuromorphic circuit design, the dynamics of each node go beyond spiking neurons by exploiting the spatiotemporal structure of the light pulses as well as the rich nonlinear dynamics of the nanophotonic units.

Speaker: Irene Fagnani

Presentation_Mode_workshop_Fagnani_Irene_EU.pdf

Presentation_Mode_workshop_Fagnani_Irene_EU.pptx

10 Accelerating MadGraph: challenges and updates

The High-Luminosity LHC (HL-LHC) upgrade presents new challenges to the computing infrastructure of the LHC experiments. Monte Carlo event generation is projected to account for 10–20% of total CPU usage at ATLAS and CMS, making the speed-ups of these tools a more pressing requirement.
To address this challenge, Madgraph5_aMC@NLO (with the MadMatrix plugin) developed an event-generation workflow with hardware acceleration in mind, including several algorithmic improvements to offload calculations of the matrix element computation on both GPUs and CPUs with vector instructions, that are increasingly populating HPC sites and grid resources.
In this seminar, I will present the performance gains obtained in typical experimental workflows, and the main challenges faced during the development.
The second part of the talk will be dedicated to MadGraph7, the next major release of the framework, and its main features, and how they combine to create a generation framework ready for the computational demands of the HL-LHC.

Speaker: Daniele Massaro (CERN)

Massaro_TALK_Sixth_MODE_Workshop.pdf

Hackathon

11 Hackathon: introduction and first steps

Speakers: Pablo Martinez Ruiz Del Arbol (Universidad de Cantabria and CSIC (ES)), Ruben Lopez Ruiz (Universidad de Cantabria and CSIC (ES))

Hackaton-MODE_v2.pdf

Free time

20:00 Dinner at OAC (included in the fee)

Thursday 3 September

Breakfast at OAC (only for people with OAC accommodation)

Applications in Medical Physics and Other Applications

Convener: Lorenzo Arsini ("Sapienza" University of Rome)

12 Towards Fully Automated Radiotherapy Planning through Differentiable Dose Models and Hybrid Optimisation

Automated radiotherapy treatment planning remains an open challenge. Current optimisation workflows typically require substantial human intervention to select beam configurations, define planning strategies, and iteratively balance competing clinical objectives. Moreover, many relevant treatment parameters are either optimised separately or selected using heuristic or experience-driven procedures.

Deep-learning dose engines offer a promising route towards fully automated planning. By learning accurate surrogate models from Monte Carlo simulations, they can provide dose estimates with Monte Carlo-comparable accuracy at a fraction of the computational cost. Crucially, these models can be made differentiable with respect to treatment parameters, allowing gradients to be propagated directly from clinical objectives to the variables defining the treatment plan.

In this talk, we will present a differentiable treatment-planning pipeline built around deep-learning surrogate dose models trained on Monte Carlo simulations. The approach combines fast dose prediction, differentiable geometric transformations, and a hybrid optimisation strategy based on the covariance matrix and gradient descent. This enables the joint optimisation of fluences, beam energies, field orientations, and other treatment parameters with minimal manual intervention.

The capabilities of the proposed approach will be illustrated through a series of increasingly realistic examples, ranging from a simplified toy problem to electron- and proton-therapy planning studies on anthropomorphic phantoms.

The framework also enables the joint optimisation of treatment plans and beam-delivery hardware. We will present an initial co-design application in which an absorber and beam collimators are optimised simultaneously with the treatment parameters. The optimisation accounts not only for the resulting therapeutic dose distribution but also for the secondary neutrons generated by interactions between the beam and these components.

Beyond nominal plan optimisation, the same framework could support robust planning by incorporating voxel-wise dose uncertainty into the objective function. It could also quantify and propagate epistemic uncertainty arising from alternative relative biological effectiveness models, enable multicriteria optimisation of competing clinical objectives, and facilitate automatic replanning in response to anatomical or treatment-related changes.

Using radiotherapy as a case study, this contribution will discuss how differentiable programming, combined with global and gradient-based optimisation, may provide the foundation for more autonomous, uncertainty-aware, adaptive, and jointly designed treatment systems.

Speaker: Carlo Mancini Terracciano

13 Differentiable simulation and optimization of superconducting quantum sensors

QSOpt (Quantum Sensing Optimization) is an end-to-end differentiable simulation and machine-learning optimization framework for open quantum networks composed of superconducting qubits, bosonic modes and input-output channels with user-defined interactions. The advancements in quantum technologies have sparked interest in employing quantum systems as sensors, with superconducting quantum circuits emerging as a flexible tool for metrology and quantum sensing, with application in fundamental physics like light dark-matter and axions detection. A promising architecture employs networks of superconducting qubits coupled to microwave bosonic modes. However, accurately modeling and optimizing these systems under realistic noise conditions requires numerical methods beyond tractable analytical approaches.
QSOpt integrates the QuTiP quantum simulation library with a JAX backend, enabling gradient-based optimization of parametrized quantum circuits for state preparation and readout, together with sweeps over hardware parameters such as couplings and dispersive shifts. This enables systematic exploration of sensing strategies and maximization of sensing performance in multi-qubit architectures under realistic conditions, while facilitating the discovery of previously unexplored sensing protocols.

Speaker: Nathan Campioni (La Sapienza, Università degli studi di Roma)

Presentazione_QSOpt_MODE2026.pdf

14 Optimizing ground-based gravitational wave detector design through differentiable programming and reinforcement learning

Over the last decades, the race for achieving a deeper understanding through more precise measurements has led physics experiments to grow increasingly complex, making their design and operation a superhuman task, even for large collaborations. Optimizing is one of the greatest hurdles of the design process, given the large dimensionality of the space of possible choices for geometry, detection technology, materials, data-acquisition, and information-extraction techniques, and the interdependence of the related parameters.

We build on techniques presented in [1] by one of the authors and apply them to the emerging field of gravitational wave science by optimizing a simulated LIGO-like interferometer [2]. In order to focus on the potential of the methodology as opposed to proposal-ready designs, the optimization is restricted to a physically relevant subspace of parameters to demonstrate key concepts such as constrained optimization with hard, physical boundaries. To motivate an upper limit to the length of the detector, reinforcement learning is carried out on a simulacrum of a plausible detector site based on real world data to find ideal detector configurations given geographical and logistical constraints.

References
[1] T. Dorigo, A. Giammanco, P. Vischia et al. (2023), Review in Physics, 10, 100085, https://www.sciencedirect.com/science/article/pii/S2405428323000047
[2] The LIGO Scientific Collaboration (2015), Class. Quantum Grav. 32, 074001, https://iopscience.iop.org/article/10.1088/0264-9381/32/7/074001

Speaker: Daniel Lanchares (Universidad de Oviedo - ICTEA)

Daniel Lanchares - MODE 2026 Slides (Google Slides link)

Daniel Lanchares - MODE 2026 Slides.pdf

10:30 Coffee break (included in the fee)

Methods and Tools

Convener: Tobias Kortus (RPTU University Kaiserslautern-Landau)

15 Spiking Neural Networks in hls4ml and their applications in physics

Spiking Neural Networks (SNNs) replace continuous activations with stateful neurons that integrate inputs over time and emit binary spikes when a threshold is crossed. This makes time part of the computation. I will present new SNN support in hls4ml, which translates models trained with PyTorch and snnTorch into clock-driven FPGA firmware. The implementation adds integrate-and-fire and leaky integrate-and-fire neurons, persistent membrane state, fixed-window execution, and configurable spike-count or membrane-potential readout, while retaining the standard hls4ml workflow. I will discuss the next steps in this work, and applications in anomaly detection for high-energy physics and beyond.

Speaker: Barry Dillon (Ulster University)

snn_hls4ml_talk_v1.pdf

16 Trust, then scale: checkable gradients in a minimal end-to-end pipeline

End-to-end differentiability is changing how we think about analysis and
experiment design: instead of scanning parameters by brute force and tuning calibrations by hand, we can let gradients of the likelihood flow back through reconstruction, digitization, and detector response, and optimize calibrations, selections, and eventually design for the physics objective directly. However, at the scale of a full simulation-and-reconstruction stack it is hard to tell whether a given differentiable method genuinely works or only appears to: it requires a lot of work upfront, ground truth is unavailable, brute-force baselines are unaffordable, and the failure modes hide.

Our prototype deliberately reduces complexity. We have built a minimal end-to-end pipeline (generation, showering, digitization, reconstruction, likelihood) small enough that we can check every gradient by hand, yet arranged so each stage maps one-to-one onto a real subsystem, so the integration path survives. Part of the motivation is pedagogical: the coupling between calibration and physics inference is far easier to see when the whole chain fits on a screen. The stronger reason is research discipline: a small system is where we can obtain ground truth, afford the brute-force baseline, and actually prove that a method is right rather than merely plausible.

The talk is organized around what this testbed establishes, and what it does not. Where the downstream is analytic, the adjoint matches finite differences to machine precision, a coding check, not a discovery, but a necessary one. When a parameter lives inside a stochastic simulator we treat as a black box, we extract a frozen local response operator from fixed-seed runs; two independent extractions agree to well under a percent, and we can measure the narrow window in which that operator still tracks a brute force rerun, rather than assume it does. Within that window a joint fit recovers a parameter buried inside the simulator alongside a calibration gain, to a few percent, where the usual fix-the-nuisance-and-scan cannot reach the buried parameter at all. It does so without re-simulating at every step, taking its gradients from cached events and the frozen operator.

We are explicit about scope: this is a closed-world injection-recovery study with detector noise switched off, the operator is only locally valid and must be refreshed, and the efficiency is structural rather than a measured speedup on anything real. What we think it earns is a small, checkable object with which to reason about end-to-end optimization before paying the cost of scale, and a shared substrate for exactly the cross-domain conversation MODE convenes. We would welcome feedback from detector, reconstruction, and differentiable-programming colleagues on which of these results survive contact with their systems, and which break first.

Speaker: Vassil Vasilev (Princeton University (US))

Code

MinLHC.pdf

Hackathon

Conveners: Pablo Martinez Ruiz Del Arbol (Universidad de Cantabria and CSIC (ES)), Ruben Lopez Ruiz (Universidad de Cantabria and CSIC (ES))

17 Hackathon session 1

Speakers: Pablo Martinez Ruiz Del Arbol (Universidad de Cantabria and CSIC (ES)), Ruben Lopez Ruiz (Universidad de Cantabria and CSIC (ES))

13:00 Lunch at OAC (included in the fee)

Free time

15:30 Coffee break (included in the fee)

Applications in Astro-HEP and Neutrino Physics

18 Optical Neutrino Telescope Design Optimization

The geometrical layout of any experiment or detector can have a large impact on its ability to produce meaningful outcomes for physics. Oftentimes we see that optimal geometries can be unintuitive. Studying and optimizing this is therefore essential. This has become a relevant topic for the optimization of cubic-kilometer-scale neutrino optical telescopes that are yet to be built. In particular, we focus on the Pacific Ocean Neutrino Experiment (P-ONE), which will have an order of 100 lines across multiple kilometers of seafloor. Set to be fully constructed in the coming years, the P-ONE geometry is yet to be finalized and studies on how to place these lines can inform crucial design decisions. Both optimal physics outcomes and engineering constraints need to be jointly considered for this. In this talk, the steps taken for geometric optimization of such an experiment will be discussed, in particular, how it will employ machine-learning techniques to apply end-to-end optimization.

Speaker: Mr Kristian Tchiorniy (Technical University of Munich)

geom_opt_nfold.gif

Mode_2026_PONE.pdf

19 A Forward-Folding Analysis Head for Neutrino Telescope Design Optimization

Optimizing the design of a neutrino telescope requires a cost function that quantifies detector performance. Generic proxies such as angular resolution or background rejection are convenient, but do not necessarily track the sensitivity of a specific analysis, which depends on the physics goal, be it source discovery or a flux measurement.
The natural cost function is therefore the sensitivity of the final analysis itself. To serve as an optimization target, however, the analysis must be fully differentiable, so that gradients can propagate through it back to the detector design parameters.
This contribution presents such a differentiable, forward-folding analysis head. Operating on Monte Carlo events from an upstream, optimizable detector simulation, it uses precalculated event weights and their gradients to efficiently compute the Fisher information, yielding a differentiable estimate of the analysis sensitivity that serves directly as the optimization objective.

Speaker: Oliver Janik

MODEPoster2026.pdf

ModeTalk2026.pdf

optimisation_freestring.gif

20 Detector-ML co-design for tau-neutrino identification in a silicon pixel–tungsten detector

We present an end-to-end optimization pipeline that designs the geometry of a tungsten-based silicon pixel detector to maximize a downstream physics objective: the separation of tau-neutrino charged-current interactions from electron- and muon-neutrino backgrounds. For candidate detector designs, a deep-learning classifier is trained on the detector hits, and a surrogate model over the design–performance relation proposes the next candidates.

The classifier is built for the co-design setting itself: a modular architecture with a shared hit encoder, per-layer summary tokens, and an inter-layer transformer, so that a single pretrained backbone transfers across detector geometries. Because generating training data dominates the cost, the pretrained backbone is fine-tuned on only a small number of events per candidate design, keeping the optimization loop cheap.

Beyond computational cost, we discuss the broader challenges of this kind of co-design: ensuring a fair comparison between detector designs, so that the optimization reflects the underlying physics rather than choices of pretraining, architecture, or training setup, and defining a physics objective that captures what the full detector system can actually measure. We present the current status of the optimization.

Speaker: Vincent Riechers (Universite de Geneve (CH))

talk_riechers.pdf

Wine Tasting and Poster Session

Lyrical gala

21 Lyrical gala

Speakers: Kalliopi Petrou (Artist), Tommaso Dorigo (INFN Padova, Luleå University of Technology, MODE Collaboration, Universal Scientific Education and Research Network)

Kolymbari_Lyrical Gala 2026 Petrou-Dorigo_MODE Workshop.pdf

20:00 Dinner at OAC (included in the fee)

Friday 4 September

Breakfast at OAC (only for people with OAC accommodation)

Hackathon

22 Hackathon session 2

Speakers: Pablo Martinez Ruiz Del Arbol (Universidad de Cantabria and CSIC (ES)), Ruben Lopez Ruiz (Universidad de Cantabria and CSIC (ES))

10:30 Coffee break (included in the fee)

Applications in Muon Tomography

Conveners: Zahraa Daher, Zahraa Zaher (Université catholique de Louvain (BE))

23 Regularization and Prior Information for Gradient-Descent-Based Reconstruction in Muon Scattering Tomography

Muon scattering tomography (MST) is a non-invasive imaging technique that utilizes the naturally occurring cosmic-ray muon flux. Although MST is now a well-established method, further improvements, particularly in image reconstruction, are necessary to fully exploit its potential. One promising approach is the use of statistical models for image reconstruction, in which a likelihood function quantifies the agreement between the voxel density map of the reconstructed object and the observed muon scattering data. In the reconstruction algorithm used in this work, this likelihood is maximized using a gradient-descent method, with gradients computed through automatic differentiation in PyTorch.

In this work, we present reconstruction results obtained with the gradient-descent method for security-related scenarios based on realistic simulation data. Furthermore, we investigate different regularization techniques and assess both their qualitative and quantitative influence on the reconstruction quality. Finally, we explore approaches for incorporating prior information into the algorithm, including material properties and geometric constraints. Such information is often available in practical applications and could therefore improve convergence behavior as well as the achievable reconstruction resolution.

Speaker: Jean-Marco Alameddine

alameddine_mode2026.pdf

24 When Synthetic Cargo Meets Real Muons: Transferring Anomaly Detection from Simulation to Operational Container Inspection

Deep learning-based anomaly detection in cargo containers using muon tomography is limited by the scarcity of anomalous data samples and the domain gap between simulation and real detector data. This research proposes a methodology to transfer anomaly detection from simulation to operational cargo container inspection. To this end, we exploit neural density estimators to approximate the distribution of benign cargo and sample physically plausible scans at scale. This generative capability drives a data augmentation strategy that combines simulated scenarios with real examples from dedicated measurement campaigns, enriching the training data with configurations that neither source alone can provide. Models pre-trained on the augmented samples are then adapted using real muon data acquisitions, allowing the learned description of normal cargo to incorporate detector responses, calibration effects, and structural variability that are absent from simulation. Anomalies emerge as statistically significant departures from this data-corrected baseline, with no prior specification of anomalous materials. We demonstrate an end-to-end deployment, from simulated scene generation to real container scans, and highlight the simulation-to-data mismatch, decisive for reliable anomaly detection.

Speaker: Angel Bueno

MODE2026Angel.pdf

25 Human phantom muon imaging simulations

Muon scattering tomography uses the natural cosmic-ray muon flux to probe dense structures without delivering ionizing dose, but its performance in low-contrast biological media remains poorly quantified. We present a simulation-based feasibility study of cosmic-ray muon tomography for medical imaging, with emphasis on pediatric spine monitoring, and connect it to human-phantom results showing that cosmic muons can measure bulk anatomical changes over time. Using GEANT4 simulation library and voxelized reference phantoms, we model controlled scoliosis deformations and evaluate whether reconstructed scattering data can recover clinically relevant spinal curvature metrics, particularly Cobb angle, under realistic detector and exposure constraints. The imaging exploits the modest density contrast between cortical bone and surrounding soft tissue. The simulation results allow us to quantify reconstruction quality as function of acquisition time.

Speaker: Konstantin Borozdin

Medical-MODE26-KBor-Sep4.pdf

26 Differentiable Large-Scale Muon Tomography: A Gran Sasso LVD Case Study (public link, slides are available only to participants in the protected contribution after login)

Speaker: Roland Grinis

13:00 Lunch at OAC (included in the fee)

Free time

15:30 Coffee break (included in the fee)

Free time

20:00 Dinner at OAC (included in the fee)

Saturday 5 September

Breakfast at OAC (only for people with OAC accommodation)

Applications in Astro-HEP and Neutrino Physics

27 Differentiable optimization pipeline for in-ice radio neutrino detectors

In-ice radio detection is a rapidly advancing field, aiming to observe the first ultra-high-energy neutrinos within the coming years. With several experiments under construction and in the planning stages, like RNO-G and the IceCube-Gen2 radio array, it is crucial to systematically explore potential detector designs to maximize the performance of the experiments and their scientific impact.
We present the first fully differentiable end-to-end optimization pipeline for in-ice radio neutrino detectors targeting ultra-high-energy neutrino observations. The framework combines differentiable PyTorch implementations of radio signal generation, propagation, detection, and reconstruction using machine-learning-based surrogate models and uncertainty estimation through the Fisher information. This enables direct optimization of detector performance metrics with respect to antenna positions and orientations. In this contribution, we focus on the development of the pipeline, how the individual components were made differentiable, and show the first proof-of-concept optimizations aimed at improving reconstruction precision for the IceCube-Gen2 radio array.

Speaker: Martin Ravn

radio_neutrino_detector_optimization_2026-09-05.pdf

radio_neutrino_detector_optimization_2026-09-05.pptx

station_layout_3d_uncertainty.gif

28 End-to-end array layout optimization for the TAMBO experiment

Abstract: Optimizing the detector array layout for the TAMBO experiment across the Colca Valley walls poses a high-dimensional placement problem intractable through conventional simulation methods. To address this, we propose an end-to-end differentiable optimization pipeline for the array layout. Built on a flow-matching shower surrogate trained on CORSIKA simulations, the pipeline chains a learned detector response network and a reconstruction neural network, enabling gradient-based position updates via backpropagation through the full surrogate chain. Trained on primary events across different detector configurations, multiple optimization chains are employed to explore the layout space and escape gradient stagnation. This represents a fully differentiable approach to neutrino telescope array placement, significantly reducing the computational cost of layout evaluation and exploring the path toward automated detector design in astroparticle physics.

Speaker: Izan Florez

TAMBO_detector_optimization_talk_IZAN_FLOREZ.pdf

TAMBO_detector_optimization_talk_IZAN_FLOREZ.pptx

29 Optimized Antenna Layouts for In-Ice Radio Neutrino Detectors

Ultra-high-energy (EeV-scale) neutrinos provide a unique probe of the most energetic astrophysical accelerators and of particle interactions at energies far beyond those accessible in terrestrial experiments. Detecting these neutrinos requires instrumenting enormous target volumes, making in-ice radio arrays a promising approach due to the long attenuation length of radio signals in glacial ice. As several next-generation radio observatories are being deployed or planned, optimizing detector layouts is essential for maximizing their scientific reach.
In this contribution, we present optimized detector layouts for in-ice radio stations obtained using a novel, fully differentiable end-to-end optimization pipeline that improves antenna positions and orientations with respect to a physical performance metric. We discuss the resulting design trends and compare their reconstruction performance with existing layouts.

Speaker: Nicolai Weitkemper

MODE26_Weitkemper_Optimized_Layouts.html

MODE26_Weitkemper_Optimized_Layouts.pdf

10:30 Coffee break (included in the fee)

Methods and Tools

30 Scaling Out hls4ml for ASICs: A Surrogate Model for\\ Neural Network Accelerator Synthesis Metrics

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 synthesis runs that can take hours per design. This not only makes exhaustive design-space exploration infeasible but also limits emerging AI-driven and agentic design flows, in which an optimization loop may need to evaluate thousands of candidate designs.

A surrogate model sidesteps the synthesis bottleneck by learning to predict synthesis metrics directly from a design specification and synthesis directives, without running synthesis at all.

Such models were recently introduced for hls4ml accelerators on FPGA targets, but two challenges remain: extending the approach to ASIC synthesis and generalizing predictions to architectures outside the training distribution. We address both.

We present a large-scale ASIC synthesis dataset and surrogate models that predict area, latency, and throughput for dense neural network accelerators generated by hls4ml and synthesized with Siemens Catapult HLS.

The dataset contains over half a million designs targeting Nangate $45\,\mathrm{nm}$, spanning a wide range of network depths, layer widths, bitwidths, and reuse factors, on which we train and compare Transformer and Graph Neural Network surrogates.

On held-out designs, our best model achieves $R^2 = 0.9997$ for latency and $R^2 = 0.9991$ for area; throughput is computed exactly rather than learned. These results, however, reflect interpolation within the training distribution.

The second challenge surfaces under depth extrapolation: when trained on $n$-layer designs and tested on deeper $(n+m)$-layer networks, the same models degrade sharply; at $n=2$ and $m=1$, for example, latency $R^2$ falls to $0.634$ and area $R^2$ to $0.789$.

The degradation suggests that, although the models interpolate accurately, they do not fully capture the structural overheads that each additional layer introduces, including control logic, pipeline coordination, and intermediate buffering.

Integrating a sum-decomposition strategy into the model restores near-interpolation accuracy, with latency $R^2$ of $0.9954$ and area $R^2$ of $0.9797$.

Finally, we demonstrate cross-node transfer: a model pre-trained with Nangate $45\,\mathrm{nm}$ fine-tunes to the room-temperature GlobalFoundries 22FDX PDK using only $10\%$ of the pre-training data volume (${\approx}50{,}000$ designs), reaching latency $R^2$ of $0.9984$ and area $R^2$ to $0.9393$.

This establishes open-library pre-training as a scalable data strategy for nodes whose PDKs are not yet ready for large synthesis campaigns, such as the cryogenic extension of 22FDX targeted for in-cryostat readout, and reduces the cost of evaluating a candidate design from hours of synthesis to milliseconds of inference, fast enough to embed in automated design-space exploration and agentic optimization loops.

Speaker: Miaoyuan Liu (Purdue University (US))

scaling-out-hls4ml-for-asics-editable.pdf

31 Exploring the Boundaries of Differentiable Radiation Transport and Detector Simulation

Speaker: Jeffrey Krupa (SLAC)

Krupa_MODE_2026 .pdf

Applications in Muon Tomography

32 3D Scene Reconstruction Across Different Physical Modalities with Probabilistic Particle Fields

Inverse problems such as computed tomography, optical inverse rendering, thermal imaging, and muon tomography arise in a wide range of scientific, medical, and security applications and are usually solved with highly specialized algorithms. By approaching these problems from a physical perspective and reformulating them in terms of particle transport and interactions, we formulate a unified forward model that maps unknown scene parameters to measurements across particle-mediated modalities.

Modeling the scene as a probabilistic particle field allows us to formulate 3D reconstruction as a maximum likelihood estimation problem with a set of differentiable optimization losses. The framework consistently captures emission-absorption, refraction, and limited multiple scattering while remaining independent of the underlying sensing modality. Measurements from different modalities, including hyperspectral, visible and infrared imaging as well as muon scattering tomography, can be combined into a joint likelihood, improving reconstruction stability and reducing ambiguities through complementary information.

We validate our method using simulated and real data acquired from optical, thermal and hyperspectral sensing systems. To demonstrate the generality of our framework, we also incorporate simulated and experimental data from muon scattering tomography to enable physically consistent 3D scene reconstruction across diverse sensing modalities.

Speaker: Felix Sattler (Detusches Zentrum für Luft- und Raumfahrt e.V. (German Aerospace Center))

sattler_mode_05_09_2026_1056.pdf

33 Machine Learning Techniques to Enable Fast Muon Scattering Tomography Measurements in a High-Rate Gamma Environment

Condition monitoring and inspection (CM&I) of nuclear wastes is required to demonstrate that wastes evolve as expected and can be safely stored in the long term. Muon scattering tomography, performed using scintillating fibre tracking detectors, has been identified as a viable approach to perform non-destructive CM&I.
However, the monitoring of nuclear wastes can be difficult due to the high rate of gamma rays that are produced through radioactive decays within the waste. This makes the use of a scintillating fibre tracker difficult for the detection of cosmic-ray muons, where every muon hit may be accompanied by thousands to tens of thousands of gamma hits.
To overcome a dominant gamma signal, Graph Neural Networks (GNNs) can be deployed to perform track identification of coincident muon hits to construct tracks. To optimise imaging times, Convolutional Neural Networks (CNNs) can be used in volumetric-post processing to identify the minimum total events (and therefore imaging time) needed to perform two-sigma confidence measurements. In this simulation-based study, GNNs yield a 97% useful muon efficiency and a 98% gamma noise rejection, while CNNs demonstrate that two-sigma confidence measurements can be reached using only 60–75% of the muon events previously thought necessary - reducing required imaging time accordingly.

Speaker: William O’Donnell

MODE_2026_WilliamO_v3.pdf

13:00 Lunch at OAC (included in the fee)

Free time

15:30 Coffee break (included in the fee)

Applications in Particle Physics

34 A Differentiable Model of Scintillation Light Collection in a Neuromorphic Readout Chain

In the context of the PHINDER project (Picosecond-scale Photonic Heterogeneous Integrated Neuromorphic Detector), we present a differentiable model describing the generation of scintillation light and its propagation, collection, and focusing onto nanowires, which constitute the sensing elements of a neuromorphic readout system. The model is designed to reproduce the main optical processes involved in the detector while preserving differentiability throughout the simulation chain. This feature allows its integration into gradient-based optimization framework, enabling the study and optimization of detector and readout parameters within a unified computational approach.

Speaker: Lorenzo Pepa

MODE_Workshop_2026.pdf

35 Towards the differentiable design of a neutron tomography system.

We present the optimization of a neutron tomography system based on a stack of Parallel-Plate Avalanche Counters with Optical Readout (O-PPACs). Building on previous work that optimized the design of a single O-PPAC using differentiable programming, we extend this framework to the complete detector stack, covering both simulation and reconstruction within the differentiable pipeline. We employ generative models as differentiable surrogates for the Geant4 detector simulation, while position reconstruction is performed by a neural network based model that outperforms traditional methods. With this framework, we aim to jointly optimize detector configuration, cost, efficiency, and spatial precision, demonstrating the potential of differentiable programming for the design of imaging systems in neutron tomography.

Speaker: Maria Pereira Martinez (Universidade de Santiago de Compostela (ES))

NeutronTomographyCrete.pdf

NeutronTomographyCrete.pptx

36 Towards automatic optimization of fast detector simulators for high-energy physics

Simulating the detector response of collider events is a computationally expensive task that involves optimizing thousands of parameters to best mimic real-world data. For parametrized fast simulators, this reduces to fitting the coefficients of predefined smearing functions, a task well-suited to gradient-based optimization via automatic differentiation. Using Delphes3 as a case study, I will discuss how parametrized detector simulators can be implemented to achieve differentiability with respect to detector parameters, thereby enabling gradient-based optimization over the entire detector configuration space.

Speaker: Luigi Favaro (Universite Catholique de Louvain (UCL) (BE))

diff_delphes.pdf

37 Tutorial: how to use LLMs via API

Speaker: Gordon Watts (University of Washington (US))

2026-09-05 - API Tutorial.pdf

2026-09-05 - API Tutorial.pptx

Github Repo

key.txt

Hackathon

38 Hackathon session 3

Speakers: Pablo Martinez Ruiz Del Arbol (Universidad de Cantabria and CSIC (ES)), Ruben Lopez Ruiz (Universidad de Cantabria and CSIC (ES))

Hackaton-MODE_closing.pdf

19:00 Free time

20:00 Gala dinner and Cretan dances (at OAC, included in the fee)

Sunday 6 September

08:00 Breakfast at OAC (only for people with OAC accommodation)

Methods and Tools

39 BRICKS: Compositional Neural Markov Kernels for Zero-Shot Radiation-Matter Simulation

We introduce a new strategy for compositional neural surrogates for radiation-matter interactions, a key task spanning domains from particle physics through nuclear and space engineering to medical physics. Exploiting the locality and the Markov nature of particle interactions, we create a \emph{next-particle prediction} kernel using hybrid discrete-continuous transformer models based on Riemannian Flow Matching on product manifolds. The model generates variable-sized typed sets of particles and radiation side effects that are the result of the interaction of an incident particle with a material volume. The resulting kernel can be composed to simulate unseen large-scale material distributions in a zero-shot manner. Unlike mechanistic simulators, our model is designed to be differentiable, provides tractable likelihoods for future downstream applications. A significant computational speed-up on GPU compared to CPU-bound mechanistic simulation is observed for single-kernel execution. We evaluate the model at the kernel level and demonstrate predictive stability over multi-round autoregressive rollouts. We additionally release a novel 20M-event radiation-matter interaction dataset for further research.

Speaker: Richard Hildebrandt

fm_cube_anim.gif

lego_mode_talk.pdf

prior_grid_electron.gif

40 Biology-inspired Model Building

Presented here is a method for model discovery and intelligent sampling of physical systems that uses recursive generation of cellular modules combined with Ant Colony Optimization (ACO). The program begins with a random sample of data and a single computational module, somewhat analogous to a single cell. This module contains several subcomponents, called organelles, that perform simple mathematical or logical operations. These organelles can themselves create submodules, allowing for the generation of recursive modules and arbitrarily complex functions. During the forward pass, the program optimizes the organelle functions using an ACO algorithm, while dual-annealing (simulated annealing combined with gradient descent) is used to optimize the model parameters during the backward pass. Early results indicate that dual-annealing performs efficiently even in high dimensions, and in contrast to other optimization methods, ACO is parallelizable, can run continuously, and adapts to changes in the graph in real-time. The ability to perform intelligent sampling falls out of this configuration naturally due to the probabilistic structure of the module. Specifically, the sampling policy is a form of unsupervised reward-free exploration-by-disagreement. The module's output distribution is used to identify the regions of highest variation among candidate models, which are then sampled for optimum information gain. As a test case, this method is applied to the discovery of interatomic potentials in atomic and molecular simulations.

Speaker: Stephen Casey (University of Miami)

Biology-Inspired Model Building - MODE 2026.pptx

Google drive slides

H2_sampling_4k.mp4

mcts_animation.gif

41 Efficient Global Path Planning in Multi-Scale Environments Using an Adaptive Rapidly-Exploring Random Tree Algorithm

Autonomous vehicles increasingly operate in large-scale environments that combine wide-open regions with confined, cluttered areas. This is exemplified by maritime missions for monitoring and inspecting critical infrastructure. Such applications require global path planning methods that can handle environments with strongly varying spatial scales. Narrow passages require fine-grained exploration, while large open regions benefit from larger planning steps to avoid inefficient search.

Existing methods for global path planning usually make use of sampling-based algorithms to avoid the discretization of large continuous planning spaces. However, these methods often use fixed expansion step sizes or heuristics tuned to specific environments. This makes them difficult to parameterize when the same planner has to operate across regions with different scales.

In this work, we present an adaptive variant of the Rapidly‑Exploring Random Tree (RRT*) path planning algorithm for multi-scale environments, with a particular application in the maritime domain. The proposed method dynamically adjusts the step size during planning based on expansion failures, enabling the planner to adapt to local constraints. This provides a simple mechanism for improving planning behavior across both constrained and open regions, without environment-specific tuning.

We evaluate the approach using benchmark scenarios derived from real coastlines and port infrastructure extracted from OpenStreetMap data. The results show improved planning performance compared with fixed-step configurations and demonstrate the benefit of adapting to local conditions in heterogeneous environments.

The presented work provides a sampling-based global planning baseline, demonstrated in maritime environments. As a next step, we aim to investigate reinforcement learning-based extensions that handle partial observability and dynamic effects such as environmental disturbances and moving obstacles.

Speaker: Robin Brase

brase_mode2026.pdf

10:30 Coffee break (included in the fee)

Applications in Particle Physics

42 Online Tracking Challenges and GNN-Based Solutions for the STCF High-Level Trigger

The Super Tau-Charm Facility (STCF) is a next-generation electron-positron collider proposed by China. The STCF will operate in the center-of-mass energy range of 2 to 7 GeV, with a design luminosity of 0.5×10³⁵ cm⁻²s⁻¹ at 4 GeV. While the high luminosity significantly enhances its physics potential, it also introduces considerable challenges, including elevated background levels and large data volumes. Background simulations indicate a data size of 35 kB per event, which is notably high for a GeV-scale electron-positron collider. Meanwhile, the drift chamber, which is the core tracking detector of the STCF, exhibits a background occupancy of approximately 8% under normal conditions. With a safety factor of 3 applied in the simulations, the projected worst-case occupancy could reach around 20%, posing a serious challenge. Conventional online and offline tracking algorithms for drift chambers typically require occupancy below 15%. In the STCF high-level trigger, the required computing latency is roughly 100 ms, making the trade-off between online tracking efficiency and latency highly precarious for classical tracking methods. To address this, approaches based on neural networks, such as GNNs, are being explored for background hit filtering and for achieving high online tracking efficiency under stringent latency constraints. Future work will focus on developing more robust and versatile GNN architectures, as well as investigating end-to-end online tracking for the STCF high-level trigger.

Speaker: Zhujun Fang

43 Machine Learning Methods for the CMS HGCAL Reconstruction Chain

The High-Luminosity LHC (HL-LHC) will significantly increase event rates and pileup, requiring reconstruction algorithms that remain efficient, scalable and robust in highly populated detector environments. To address this challenge, CMS will replace its endcap calorimeters with the High Granularity Calorimeter (HGCAL), a highly segmented calorimeter providing fine three-dimensional spatial information, energy measurements and precision timing in the forward region. CMS is developing The Iterative CLustering reconstruction framework (TICL) within the central CMS software (CMSSW) to reconstruct particles in HGCAL, progressing from calibrated detector hits to layer clusters, three-dimensional tracksters and final particle-flow candidates. Machine learning (ML) plays a key role in this chain, including particle identification, electromagnetic superclustering, hadron energy regression and the linking of reconstructed components into particle candidates. This talk summarizes current ML applications in HGCAL event reconstruction, with emphasis on how detector granularity, shower topology, timing and track compatibility are used to guide reconstruction decisions. Recent studies on ML driven electromagnetic reconstruction are highlighted, including a learning-based approach for associating and merging reconstructed components in ambiguous multi-candidate configurations, with timing information included for the first time in this superclustering context. Future directions toward deeper integration of precision timing and more unified learning-based reconstruction strategies are also discussed.

Speaker: Gamze Sokmen (Centre National de la Recherche Scientifique (FR))

TICL_ML_MODE_Sokmen.pdf

Applications in Medical Physics and Other Applications

44 Hybrid optics for imaging endoscopy: where co-design can happen

Miniaturized imaging probes play a critical role in the development of fiber-optic microendoscopy. Scanning fiber endoscopy techniques, incorporate compact distal scanning probes paired with micro-objective elements such as gradient refractive index (GRIN) lenses. However, systems relying solely on GRIN optics often experience off-axis aberrations, which reduce image quality particularly toward the edges of the field of view. Addressing these aberrations without enlarging the distal optical assembly remains a significant challenge in miniaturized designs. Artificially engineered metamaterials have attracted significant attention for their ability to manipulate electromagnetic responses beyond those of natural materials. The planar version of metamaterials, called metasurfaces, is a very promising technology that allows complex manipulation of light while keeping miniaturization. Metasurfaces in optical applications are often called metalenses.

We have developed hybrid solutions where a metalens is combined with a GRIN lens to enhance the optical performances while keeping a compact form factor. A hybrid GRIN-metalens was first designed for reflectance imaging, minimizing on-axis aberrations and also off-axis aberrations due to the extreme angles of the field of view. Then, this hybrid optical solution was extended towards fluorescence imaging where emission and collection wavelengths are different and other design constraints are added. Both solutions have been fabricated and validated, and lead to high-resolution endoscopic lens systems that maintains uniform image quality across the scan field. Simulation and experimental results will be presented.

Co-design of hardware and software is emerging in imaging applications, where optical hardware and computational algorithms are jointly optimized for the overall system performance. This important aspect will also be discussed during the talk.

Speaker: Francesco Ferranti

45 Differentiable Molecular Modeling: Learning Interaction Potentials from Experiment and Simulation

Machine-learning interatomic potentials offer a flexible and efficient way to represent complex potential-energy surfaces and enable molecular simulations at scales beyond those accessible to electronic-structure methods. They are commonly trained in a bottom-up manner using labeled reference data, such as energies and forces obtained from density functional theory. While such models can reproduce the underlying calculations with high accuracy, their predictions do not necessarily agree with measured observables. Deviations may arise from limitations of the reference method, limited coverage of the relevant configurational space, or simplifications inherent in the molecular representation.

Training directly on experimental observables can help correct these discrepancies. Differentiable Trajectory Reweighting (DiffTRe) combines automatic differentiation with statistical-mechanical reweighting to compute gradients of equilibrium thermodynamic, structural, and mechanical properties with respect to model parameters. Rather than differentiating through entire molecular-dynamics trajectories, DiffTRe establishes a differentiable path through a reweighted ensemble, avoiding the associated memory requirements and unstable gradients. This enables interaction potentials to be refined efficiently in a top-down manner while retaining information learned from microscopic simulations.

In our work, we have applied top-down optimization strategies to data-efficient coarse-grained model design, the training and refinement of implicit-solvent potentials, and the correction of machine-learning interatomic potentials using both simulation and experimental data. The corresponding algorithms are implemented in chemtrain, our open-source, JAX-based framework for combining top-down and bottom-up training objectives. More broadly, integrating measured observables with microscopic simulation data provides a general strategy for developing molecular models that are both physically grounded and consistent with experiment.

Speaker: Jan Eckwert (Technical University of Munich)

TWIN_JanEckwert.pdf

13:00 Lunch at OAC (included in the fee)

14:00 Free time

15:30 Coffee break (included in the fee)

Applications in Particle Physics

46 Transformer-based reconstruction of electromagnetic showers in CMS

The reconstruction of electrons and photons in the CMS Electromagnetic Calorimeter (ECAL) currently relies on a geometrical clustering algorithm called PFClustering. While it is efficient for isolated particles, it has a limited ability to resolve close-by showers and mitigate detector noise, which reduces the sensitivity of physics analyses and will worsen with detector ageing. We present ClusTEX, a single-step graph-attention-based transformer that reconstructs photon energy and position directly from calorimeter readout. Through a novel positional encoding scheme, ClusTEX remains sensitive to the relative graph structure of the events while simultaneously learning detector-position dependencies. We demonstrate that this approach eliminates the need for a dedicated classification step and remains robust to complex event topologies and non-responsive detector regions. Trained and evaluated on a toy ECAL simulation, ClusTEX outperforms both PFClustering and GNN-based approaches: for events with overlapping showers, PFClustering achieves 82.0% signal efficiency with an energy resolution of 6.24 GeV, compared to 98.6% and 0.80 GeV for ClusTEX. The impact of this improvement is reflected in the reconstruction of boosted $\pi^0\rightarrow\gamma\gamma$ decays, where ClusTEX successfully resolves pions above 30 GeV that are inaccessible to PFClustering. We report on the ongoing integration of ClusTEX into CMSSW and present first results using the full CMS simulation.

Speaker: Yuliia Maidannyk (Université Paris-Saclay (FR))

MODE_clustering.pdf

47 End-to-end calorimeter optimization using generative surrogate models

Recent advances in generative machine learning have opened new avenues for detector optimization by fully exploiting high-dimensional parameter spaces. In this work, we build on the AI Detector Optimization (AIDO) framework and extend it to include particle identification using calorimeter shower shapes, as well as more advanced exploration strategies. In addition, preliminary work involving more advanced ML-based reconstruction algorithms has demonstrated that optimization can be performed completely end-to-end, directly inferring multiple particles, their positions, energies, and types. Finally, we consider both hardware detector parameters and reconstruction software parameters to provide a complete co-design approach.

Speaker: Dr Florian Bury (University of Bristol)

MODE workshop 26 - AIDO.pdf

MODE workshop 26 - AIDO.pptx

48 CaloFound: A Geometry Standardized Foundation Model for Calorimeters

Calorimeter reconstruction and simulation are usually developed separately for each detector geometry, segmentation and physics task. This often requires dedicated algorithms for energy regression, position reconstruction, clustering, shower shape prediction, leakage correction and simulation. We present CaloFound, a foundation model approach for calorimeters based on geometry standardization and masked autoencoding.

The method natively supports simulation and can be modified to support reconstruction oriented tasks in a common framework. A geometry standardization network first maps the native detector response into a common higher dimensional representation space, preserving local detector information while reducing the dependence on the original readout geometry. A masked autoencoder is then trained on regions around seed candidates above an energy threshold. Part of the calorimeter response near the seed (deposit above certain energy threshold) is masked and the network learns to infer the missing deposits from the visible shower information and encoded shower information (center and energy). This objective allows the model to learn detector response, shower morphology, local energy correlations and shower completion directly in the standardized detector space.
The algorithm is studied using GEANT4 simulation for single layer lead tungstate, $\mathrm{PbWO_4}$, crystal based calorimeter and for sampling calorimeters consisting of lead passive absorber layers and silicon active material with different cell size configurations. In the multilayer case, the self-supervised objective is extended to predict the shower shape in the next calorimeter layer. This allows the same model to learn longitudinal shower development and to support simulation tasks such as layer-to-layer shower development, shower completion and shower-tail leakage estimation.
After pretraining, the masked-autoencoder head can be replaced by task-specific regression or classification heads. The pretrained backbone is applied to energy and position regression, cluster cell or crystal assignment, shower shape estimation, leakage prediction and simulation-related shower prediction tasks. For energy regression in sampling calorimeters, the method performs as well as LSQ-based and TF-based regression methods and as well as a dedicated GNN-based algorithm in the low noise regime. In the high noise regime, it outperforms the traditional techniques and performs similarly to the dedicated GNN-based algorithm.

With CaloFound we demonstrate that calorimeter data can be represented with a reusable foundation model that supports both reconstruction and simulation tasks. By combining detector geometry standardization with masked autoencoding, the method provides a common backbone for homogeneous and sampling calorimeters and enables efficient adaptation to multiple detector configurations, noise conditions and physics objectives.

Speaker: Ozgur Sahin (Université Paris-Saclay (FR))

CaloFound_Mode.pdf

Closing session: Closing Session

Convener: Prof. Pietro Vischia (Departamento de Física and ICTEA, Universidad de Oviedo)

49 Workshop closing and awards

Speaker: Prof. Pietro Vischia (Departamento de Física and ICTEA, Universidad de Oviedo)

2026 - Best Poster Award.pptx

Sixth MODE Workshop Closing 2026.pptx

18:00 Free time

20:00 Dinner at OAC (included in the fee)

Monday 7 September

08:45 Departure day