Pith. sign in

REVIEW 4 major objections 7 minor 34 references

Physics Instrument Design with Reinforcement Learning

T0 review · 4 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Reinforcement learning can design particle detectors from scratch, and the resulting layouts beat hand-tuned baselines in simulation.

desk verdict A genuine proof-of-concept for RL in detector design, but the calorimeter numbers depend on an unjustified fixed-shower shortcut. read the letter →

arxiv 2412.10237 v1 pith:LY2U5FAY submitted 2024-12-13 physics.ins-det cs.AIhep-ex

Shah Rukh Qasim , Patrick Owen , Nicola Serra This is my paper · ORCID
classification physics.ins-detcs.AIhep-ex PACS 29.40.Vj29.40.Gx
keywords reinforcementlearningdetectordesigncalorimetersegmentationspectrometeroptimizationproximalpolicyGeant4simulationmixedactionspacefuturecircularcollider
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Reinforcement learning, which learns by trial and error instead of following gradients, can design particle-physics detectors from scratch, placing a variable number of layers and choosing their properties one step at a time. This paper demonstrates the claim in two simulated settings: a sampling calorimeter for energy measurement and a spectrometer for momentum measurement. In the calorimeter study, the RL-designed layout improves hadronic energy resolution from 34.13% to 25.27% at 50 GeV while matching the baseline for electromagnetic showers. In the spectrometer study, the RL layout improves 10 GeV momentum resolution from 6.48% to 3.74% at 100% efficiency. If the result holds, it gives a path to optimizing instruments without predefining a parametric model, directly relevant to future projects like the FCC.

What carries the argument

The engine is a mixed-action-space episodic Markov decision process, solved with Proximal Policy Optimization (PPO). At each step the agent outputs a continuous $\Delta z$ placing the next layer or station and a discrete choice of sensor type or granularity; it observes only the current longitudinal position and the budget used, not the full layout (a partially observable setup). The environment returns zero reward until the final step, when the agent receives a design score: calorimeter resolution for 50/100 GeV photon and pion showers in the first study, and a resolution-plus-efficiency score from straight-line momentum reconstruction in the second. PPO's clipped objective keeps policy updates stable while exploration lets the agent escape local minima.

What would settle it

Resimulate each RL-proposed calorimeter layout from scratch (no deposit reuse) with sufficient statistics and compare; if the 25.27% hadronic resolution degrades to baseline levels, the claimed improvement is an artifact of fixed deposits. For the spectrometer, re-score the RL layout using an independent reconstruction algorithm; if resolution or efficiency at 10 GeV reverts to baseline values, the design is specialized to the hand-coded fitter.

Watch

Extended reading notes

Core claim

The paper's central claim is that a reinforcement learning agent, given only a budget and a terminal score, can construct particle detector layouts from scratch that outperform hand-coded baselines in simulation. In the calorimeter study, the agent sequentially places active layers with a continuous action for spacing and a discrete action for sensor type; the resulting design improves mean-corrected hadronic resolution from 34.13% to 25.27% at 50 GeV and from 24.48% to 17.79% at 100 GeV, while the electromagnetic resolution stays statistically unchanged. In the spectrometer study, the agent places tracking stations and chooses their pixel granularity, yielding 10 GeV momentum resolution of 3.74% at 100% efficiency, compared with 6.48% and 5.97% for two baselines. The authors present these results as evidence that RL's exploration avoids the local optima that plague gradient-based methods and that it can make the discrete, combinatorial decisions -- number and type of layers -- that differentiable optimization cannot.

Load-bearing premise

The simulation reuses a single set of Geant4 energy deposits to score every calorimeter design, assuming that moving or retyping the active layers does not change how the shower develops; the spectrometer scores are computed with one hand-coded line-fitting reconstruction, so the optimized station positions are optimal for that reconstruction.

Editorial extensions

If this is right

  • RL can explore design spaces that gradient-based methods cannot enter, including the number of layers, sensor type, and granularity.
  • Because the simulator does not need to be differentiable, the method applies to any Monte Carlo prediction that can be turned into a score.
  • The agent rediscovered a known design principle -- dense layers in the front for short electromagnetic showers -- without being told it.
  • The resulting designs are starting points for further fine-tuning by differentiable methods, which the authors explicitly allow.
  • The same layer-by-layer formulation can be extended to more complex instruments with graph-based reconstruction and surrogate simulators.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported gains are computed with a fixed scoring pipeline; re-running the calorimeter study with fresh Geant4 showers for each design, and the spectrometer study with an independent reconstruction algorithm, would show how much of the improvement is intrinsic to the geometry.
  • Co-optimizing a graph neural network reconstruction with the geometry, as the outlook suggests, could move the optimal station positions, since the current agent optimizes for a hand-coded line fitter.
  • The 'from scratch' claim lives inside the allowed action envelope (budget, z range, sensor menu); opening material choice and full 3D layout is the next stress test.
  • Scoring at two particle energies with 2500 events per species means the designs are fitted to those benchmarks; evaluating over continuous energy spectra is a natural next check.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes Reinforcement Learning (RL) as an alternative to gradient-based and surrogate-based methods for physics instrument design. It presents two proof-of-concept studies: longitudinal segmentation of a sampling calorimeter (Section 4.1) and combined longitudinal/transverse placement of tracking stations in a spectrometer (Section 4.2). In both cases a PPO agent with a mixed continuous/discrete action space sequentially places detector components, and the final designs are compared against hand-crafted baselines in Tables I and II. The authors claim that RL offers inherent exploration, avoids local minima, and can handle discrete, variable-number design choices without a predefined parametric model.

Significance. If the results hold, the paper would be a useful proof-of-concept that RL can explore discrete and combinatorial detector-design spaces that are inaccessible to gradient-based methods. The problem formulation is clear, the training curves and design visualizations are informative, and the qualitative finding that the agent learns physically sensible strategies (such as front-loaded calorimeter layers) is interesting. However, the current evidence does not support the strongest claims: the calorimeter evaluation reuses a fixed simulation set, no differentiable or surrogate baseline is run, and the reward functions are exactly the metrics used for final evaluation. These issues must be resolved before the paper can make its advertised contribution to the instrumentation community.

major comments (4)
  1. [Section 4.1, Eq. (2)] The performance of every candidate calorimeter design is scored by reusing a single set of Geant4 energy deposits, justified by the sentence 'Given that it is a sampling calorimeter, it allowed reuse of the same set of simulations across designs.' Since the agent's actions include selecting among three sensor types with different material budgets (0.12, 0.15, 0.20) and placing them at variable positions, this reuse is only valid if the active layers have negligible effect on shower development. No demonstration of this is provided, and for hadronic showers the assumption is strong: changing absorber/sensor materials and thicknesses can modify the shower and sampling fluctuations. Therefore the hadronic resolutions in Table I (25.27% at 50 GeV) are not established as the performance of a physically realizable detector. The authors should rerun the simulation for the final design (or several top designs) and confirm that the reported improvement persists.
  2. [Abstract, Sections 2 and 4] The paper claims that RL is an alternative to 'differentiable programming and surrogate-based differentiable design optimization methods' and that it offers 'unique advantages' over them, but no competing differentiable, surrogate-based, or Bayesian optimization baseline is evaluated anywhere in the manuscript. Tables I and II compare the RL results only to hand-crafted baseline designs. To substantiate the central contribution, the authors should include at least one standard baseline (e.g., Bayesian optimization, a differentiable surrogate, or even a random/greedy design search) and compare performance and computational cost. If such a baseline is not feasible, the abstract and conclusion should be substantially weakened to present RL as a feasible approach rather than a superior one.
  3. [Section 4.2, Eqs. (5)-(6)] The spectrometer score is computed with a custom, hand-coded line-fitting reconstruction described in detail in the bullet list under Figure 5. The optimized station positions and granularities are therefore optimal only for that particular reconstruction algorithm. If the design objective is intended to be generic tracking performance, the authors should test the RL-designed spectrometer with at least one independent reconstruction method (e.g., a simple Kalman filter or a graph-based track finder) to show that the improvement in Table II is not an artifact of the chosen fitting procedure. Alternatively, the claims should be explicitly scoped to the specific reconstruction used.
  4. [Section 4.1 and 4.2, Eqs. (2), (5)-(6)] In both studies the reward that the RL agent maximizes is the same function used to evaluate the final design in Tables I and II. This is not circular in a strict sense because the score is computed from independent simulations of test particles, but it means that the reported improvements over baselines are exactly the quantity the agent was trained to maximize. The threshold constants in Eq. (2) and the coefficients and floors in Eqs. (5)-(6) are free parameters; the paper does not report how the results change under plausible variations of these constants, nor does it justify the chosen values. The robustness of the claimed improvement to reward-shaping choices should be studied or at least discussed.
minor comments (7)
  1. [Abstract and Section 1] The abstract contains 'It's applicability' which should be 'Its applicability'; Section 1 has a typo 'Otimization' for 'Optimization'.
  2. [Figure 4 caption] The caption says 'the performance of the calorimeter' but the figure shows the spectrometer; also 'the the best design' contains a duplicated article.
  3. [Section 4.2] The text says the best design is selected over intervals of 400 designs, while the Figure 4 caption says intervals of 20000 designs; these numbers should be reconciled.
  4. [Section 4.2] The formula for sensor granularity is garbled as '500Agran + 10) · 500)'; it should be stated as (A_gran + 10) × 500 pixels.
  5. [Section 4.2] The phrase 'the stripped section' describing the magnet in Figure 4 should be 'the striped section'.
  6. [Section 5.3] The sentence 'This was as choice of convenience' should be 'This was a choice of convenience'.
  7. [General] The paper does not state whether code and data will be released. For a methods paper, providing the training code, environment details, and exact simulation configuration would greatly improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RL designs are scored by independent Geant4 simulation and compared against baselines; reporting the optimized objective is standard optimization, not a circular derivation.

full rationale

The paper's central claim is that RL can optimize detector layouts, and the evidence is the resolution and efficiency of the final design compared with baselines. These metrics are computed from Geant4 simulation (calorimeter) or from a dedicated simulator with line fitting (spectrometer), not derived algebraically from the reward function. The reward in Eq. (2) and Eqs. (5)-(7) is defined directly in terms of those same metrics, so reporting improvement on them is the normal closed-loop optimization procedure: the optimized quantity is the reported quantity, but it is evaluated independently for each candidate design. The calorimeter's reuse of a single set of energy deposits across designs is a physical modeling assumption about non-perturbing sampling layers; if false, the results would be invalid, but the assumption is not equivalent to assuming the conclusion, so it is a correctness risk rather than a circular step. Self-citations (Qasim et al. 2019, 2022) appear only for calibration details and future reconstruction proposals, not as load-bearing evidence for the RL result. No uniqueness theorem or ansatz is imported via self-citation, and no fitted parameter is renamed as a prediction. Therefore no circular step can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on a handful of hand-chosen reward parameters and on the fidelity of two simplified simulation models. No new physical entities are introduced; the 'design score' is an objective function, not an entity.

free parameters (3)
  • Calorimeter reward thresholds
    Constants 8%, 5%, 25%, 18% in Eq. 2 set the reference resolutions that zero out the reward. Chosen as 'slightly better than roughly expected in a reasonable design', so the optimized layout depends on this subjective choice.
  • Spectrometer reward weights and floors
    The 95% efficiency floor, 3% and 8% resolution floors, factor of 3 in Eqs. 5-6, and the penalties (-400+20*Ncorrect, -50) define the objective; different values change the optimized station placement.
  • Sensor type costs and budget cap = 0.12, 0.15, 0.20; 6.0
    The cost per sensor type and cumulative budget constrain which designs are possible; these are chosen by the authors rather than derived from data or physics.
assumptions (4)
  • domain assumption A single set of Geant4-simulated energy deposits can be reused to score any calorimeter design after moving active layers.
    Section 4.1: 'Given that it is a sampling calorimeter, it allowed reuse of the same set of simulations across designs.' Assumes active layers do not perturb showers; the entire fast-evaluation pipeline depends on this.
  • domain assumption The simplified spectrometer model and hand-coded line-fitting reconstruction are faithful proxies for real tracking performance.
    Section 4.2 and Figure 5: uniform 1 T field, straight-line segments, Eqs. 3-4, and the iterative fit are used to compute the optimized score. The optimized station positions are only optimal for this reconstruction.
  • domain assumption PPO with two fully-connected networks and the reduced observations can find the global optimum or a sufficiently good design.
    The paper relies on PPO's exploration to escape local minima, but provides no evidence that the policy architecture or the 2D observation space is sufficient for the task.
  • standard math Standard physics formulas (multiple scattering, momentum resolution) are valid in the regime used.
    Eqs. 3 and 4 are taken from Groom and Klein and standard HEP references; assumed correct.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics Instrument Design with Reinforcement Learning." pith.science (2026). https://pith.science/paper/LY2U5FAY

@misc{pith2026241210237,
  author       = {Pith},
  title        = {Pith review of: Physics Instrument Design with Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LY2U5FAY}},
  note         = {Machine review of arXiv:2412.10237}
}
read the original abstract

We present a case for the use of Reinforcement Learning (RL) for the design of physics instrument as an alternative to gradient-based instrument-optimization methods. It's applicability is demonstrated using two empirical studies. One is longitudinal segmentation of calorimeters and the second is both transverse segmentation as well longitudinal placement of trackers in a spectrometer. Based on these experiments, we propose an alternative approach that offers unique advantages over differentiable programming and surrogate-based differentiable design optimization methods. First, Reinforcement Learning (RL) algorithms possess inherent exploratory capabilities, which help mitigate the risk of convergence to local optima. Second, this approach eliminates the necessity of constraining the design to a predefined detector model with fixed parameters. Instead, it allows for the flexible placement of a variable number of detector components and facilitates discrete decision-making. We then discuss the road map of how this idea can be extended into designing very complex instruments. The presented study sets the stage for a novel framework in physics instrument design, offering a scalable and efficient framework that can be pivotal for future projects such as the Future Circular Collider (FCC), where most optimized detectors are essential for exploring physics at unprecedented energy scales.

Figures

Figures reproduced from arXiv: 2412.10237 by the authors.

Figure 1
Figure 1. Agent-environment interaction in Reinforcement Learning (adapted from Sutton and [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Reinforcement Learning for instrument design [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Design of uniform sampling calorimeter with proximal policy optimization (PPO). The [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Design of spectrometer with Reinforcement Learning. The top two plot shows the [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: The top left figure shows the process of line fitting in Region A and right figure, Region [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 8 canonical work pages

  1. [1]

    Gauger, Jan Kieseler, Markus Klute, and Vassil Vassilev

    Max Aehle, Xuan Tung Nguyen, Mihály Novák, Tommaso Dorigo, Nicolas R. Gauger, Jan Kieseler, Markus Klute, and Vassil Vassilev. Efficient forward-mode algorithmic derivatives of geant4, 2024. URL https://arxiv.org/abs/2407.02966

    arXiv 2024

  2. [2]

    Optimising the active muon shield for the ship experiment at cern

    A Baranov, E Burnaev, D Derkach, A Filatov, N Klyuchnikov, O Lantwin, F Ratnikov, A Ustyuzhanin, and A Zaitsev. Optimising the active muon shield for the ship experiment at cern. Journal of Physics: Conference Series, 934: 0 012050, December 2017. ISSN 1742-6596. doi:10.1088/1742-6596/934/1/012050. URL http://dx.doi.org/10.1088/1742-6596/934/1/012050

    doi:10.1088/1742-6596/934/1/012050 2017

  3. [3]

    Caloclouds: fast geometry-independent highly-granular calorimeter simulation

    Erik Buhmann, Sascha Diefenbacher, Engin Eren, Frank Gaede, Gregor Kasicezka, Anatolii Korol, William Korcari, Katja Kr\" u ger, and Peter McKeown. Caloclouds: fast geometry-independent highly-granular calorimeter simulation. Journal of Instrumentation, 18 0 (11): 0 P11025, November 2023. ISSN 1748-0221. doi:10.1088/1748-0221/18/11/p11025. URL http://dx.d...

    doi:10.1088/1748-0221/18/11/p11025 2023

  4. [4]

    Caloclouds ii: ultra-fast geometry-independent highly-granular calorimeter simulation

    Erik Buhmann, Frank Gaede, Gregor Kasieczka, Anatolii Korol, William Korcari, Katja Kr\" u ger, and Peter McKeown. Caloclouds ii: ultra-fast geometry-independent highly-granular calorimeter simulation. Journal of Instrumentation, 19 0 (04): 0 P04020, April 2024. ISSN 1748-0221. doi:10.1088/1748-0221/19/04/p04020. URL http://dx.doi.org/10.1088/1748-0221/19...

    doi:10.1088/1748-0221/19/04/p04020 2024

  5. [5]

    Gauger, Christian Glaser, Atılım G

    Tommaso Dorigo, Andrea Giammanco, Pietro Vischia, Max Aehle, Mateusz Bawaj, Alexey Boldyrev, Pablo de Castro Manzano, Denis Derkach, Julien Donini, Auralee Edelen, Federica Fanzago, Nicolas R. Gauger, Christian Glaser, Atılım G. Baydin, Lukas Heinrich, Ralf Keidel, Jan Kieseler, Claudius Krause, Maxime Lagrange, Max Lamparth, Lukas Layer, Gernot Maier, Fe...

    arXiv 2023

  6. [6]

    Electron-ion collider: The next QCD frontier: Understanding the glue that binds us all

    ePIC Collaboration . Electron-ion collider: The next QCD frontier: Understanding the glue that binds us all. The European Physical Journal A, 52 0 (9), September 2016. ISSN 1434-601X. doi:10.1140/epja/i2016-16268-9. URL http://dx.doi.org/10.1140/epja/i2016-16268-9

    doi:10.1140/epja/i2016-16268-9 2016

  7. [7]

    Fabjan and Fabiola Gianotti

    Christian W. Fabjan and Fabiola Gianotti. Calorimetry for particle physics. Reviews of Modern Physics, 75 0 (4): 0 1243–1286, October 2003. ISSN 1539-0756. doi:10.1103/revmodphys.75.1243. URL http://dx.doi.org/10.1103/RevModPhys.75.1243

    doi:10.1103/revmodphys.75.1243 2003

  8. [8]

    Catastrophic forgetting in connectionist networks

    R French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 3 0 (4): 0 128–135, April 1999. ISSN 1364-6613. doi:10.1016/s1364-6613(99)01294-2. URL http://dx.doi.org/10.1016/S1364-6613(99)01294-2

    doi:10.1016/s1364-6613(99)01294-2 1999

Show all 34 references
  1. [9]

    Geant4—a simulation toolkit

    GEANT4 Collaboration . Geant4—a simulation toolkit. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 506 0 (3): 0 250–303, July 2003. ISSN 0168-9002. doi:10.1016/s0168-9002(03)01368-8. URL http://dx...

    2003 doi

  2. [10]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014

    2014

  3. [11]

    D. E. Groom and S. R. Klein. Passage of particles through matter. The European Physical Journal C, 15 0 (1–4): 0 163–173, March 2000. ISSN 1434-6052. doi:10.1007/bf02683419. URL http://dx.doi.org/10.1007/BF02683419

    doi

  4. [12]

    Kai Guo, Zhenze Yang, Chi-Hua Yu, and Markus J. Buehler. Artificial intelligence and machine learning in design of mechanical materials. Materials Horizons, 8 0 (4): 0 1153–1172, 2021. ISSN 2051-6355. doi:10.1039/d0mh01451f. URL http://dx.doi.org/10.1039/D0MH01451F

    2021 doi

  5. [13]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840--6851. Curran Associates, Inc., 2020. URL h...

    2020

  6. [14]

    Distributed prioritized experience replay, 2018

    Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado van Hasselt, and David Silver. Distributed prioritized experience replay, 2018. URL https://arxiv.org/abs/1803.00933

    2018 arXiv

  7. [15]

    Object condensation: one-stage grid-free multi-object reconstruction in physics detectors, graph, and image data

    Jan Kieseler. Object condensation: one-stage grid-free multi-object reconstruction in physics detectors, graph, and image data. The European Physical Journal C, 80 0 (9), September 2020. ISSN 1434-6052. doi:10.1140/epjc/s10052-020-08461-2. URL http://dx.doi.org/10.1140/epjc/s1...

    2020 doi

  8. [16]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes . In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings , 2014

    2014

  9. [17]

    Le, James Laudon, Richard Ho, Roger Carpenter, and Jeff Dean

    Azalia Mirhoseini, Anna Goldie, Mustafa Yazgan, Joe Wenjie Jiang, Ebrahim Songhori, Shen Wang, Young-Joon Lee, Eric Johnson, Omkar Pathak, Azade Nova, Jiwoo Pak, Andy Tong, Kavya Srinivasa, William Hang, Emre Tuncer, Quoc V. Le, James Laudon, Richard Ho, Roger Carpenter, and J...

    2022 doi

  10. [18]

    Machine-learning optimized design of experiments (mode), 2024

    MODE Collaboration . Machine-learning optimized design of experiments (mode), 2024. URL https://mode-collaboration.github.io. Accessed: 2024-10-31

    2024

  11. [19]

    The beam and detector of the na62 experiment at cern

    NA62 Collaboration . The beam and detector of the na62 experiment at cern. Journal of Instrumentation, 12 0 (05): 0 P05025–P05025, May 2017. ISSN 1748-0221. doi:10.1088/1748-0221/12/05/p05025. URL http://dx.doi.org/10.1088/1748-0221/12/05/P05025

    2017 doi

  12. [20]

    2022 NA62 Status Report to the CERN SPSC

    NA62 Collaboration . 2022 NA62 Status Report to the CERN SPSC . Technical report, CERN, Geneva, 2022. URL https://cds.cern.ch/record/2805351

    2022

  13. [21]

    Calogan: Simulating 3d high energy particle showers in multilayer electromagnetic calorimeters with generative adversarial networks

    Michela Paganini, Luke de Oliveira, and Benjamin Nachman. Calogan: Simulating 3d high energy particle showers in multilayer electromagnetic calorimeters with generative adversarial networks. Physical Review D, 97 0 (1), January 2018. ISSN 2470-0029. doi:10.1103/physrevd.97.014...

    2018 doi

  14. [22]

    Learning representations of irregular particle-detector geometry with distance-weighted graph networks

    Shah Rukh Qasim, Jan Kieseler, Yutaro Iiyama, and Maurizio Pierini. Learning representations of irregular particle-detector geometry with distance-weighted graph networks. The European Physical Journal C, 79 0 (7), July 2019. ISSN 1434-6052. doi:10.1140/epjc/s10052-019-7113-9....

    2019 doi

  15. [23]

    End-to-end multi-particle reconstruction in high occupancy imaging calorimeters with graph neural networks

    Shah Rukh Qasim, Nadezda Chernyavskaya, Jan Kieseler, Kenneth Long, Oleksandr Viazlo, Maurizio Pierini, and Raheel Nawaz. End-to-end multi-particle reconstruction in high occupancy imaging calorimeters with graph neural networks. The European Physical Journal C, 82 0 (8), Augu...

    2022 doi

  16. [24]

    End-to-end multi-track reconstruction using graph neural networks at belle ii, 2024

    Lea Reuter, Giacomo De Pietro, Slavomira Stefkova, Torben Ferber, Valerio Bertacchi, Giulia Casarosa, Luigi Corona, Patrick Ecker, Alexander Glazov, Yubo Han, Martina Laurenza, Thomas Lueck, Ludovico Massaccesi, Suryanarayan Mondal, Bianca Scavino, Stefano Spataro, Christian W...

    2024 arXiv

  17. [25]

    Variational inference with normalizing flows

    Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530--1538. PMLR, 2015

    2015

  18. [26]

    Overview of the lhcb experiment

    Olivier Schneider. Overview of the lhcb experiment. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 446 0 (1–2): 0 213–221, May 2000. ISSN 0168-9002. doi:10.1016/s0168-9002(00)00014-0. URL http://d...

    doi

  19. [27]

    Proximal policy optimization algorithms, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. URL https://arxiv.org/abs/1707.06347

    2017 arXiv

  20. [28]

    Black-box optimization with local generative surrogates

    Sergey Shirobokov, Vladislav Belavin, Michael Kagan, Andrei Ustyuzhanin, and Atilim Gunes Baydin. Black-box optimization with local generative surrogates. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Syst...

    2020

  21. [29]

    David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madele...

    2016 doi

  22. [30]

    A general reinforcement learning algorithm that masters chess, shogi, and go through self-play

    David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, Timothy Lillicrap, Karen Simonyan, and Demis Hassabis. A general reinforcement learning algorithm that masters chess, ...

    2018 doi

  23. [31]

    Tomopt: differential optimisation for task- and constraint-aware design of particle detectors in the context of muon tomography

    Giles C Strong, Maxime Lagrange, Aitor Orio, Anna Bordignon, Florian Bury, Tommaso Dorigo, Andrea Giammanco, Mariam Heikal, Jan Kieseler, Max Lamparth, Pablo Martínez Ruíz del Árbol, Federico Nardi, Pietro Vischia, and Haitham Zaraket. Tomopt: differential optimisation for tas...

    2024 doi

  24. [32]

    Reinforcement learning: An introduction

    Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018

    2018

  25. [33]

    Particle-based fast jet simulation at the lhc with variational autoencoders

    Mary Touranakou, Nadezda Chernyavskaya, Javier Duarte, Dimitrios Gunopulos, Raghav Kansal, Breno Orzari, Maurizio Pierini, Thiago Tomei, and Jean-Roch Vlimant. Particle-based fast jet simulation at the lhc with variational autoencoders. Machine Learning: Science and Technology...

    2022 doi

  26. [34]

    Zare, and Patrick Riley

    Zhenpeng Zhou, Steven Kearnes, Li Li, Richard N. Zare, and Patrick Riley. Optimization of molecules via deep reinforcement learning. Scientific Reports, 9 0 (1), July 2019. ISSN 2045-2322. doi:10.1038/s41598-019-47148-x. URL http://dx.doi.org/10.1038/s41598-019-47148-x

    2019 doi

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.