Pith. sign in

REVIEW 3 major objections 4 minor 24 references

AI-assisted design of experiments at the frontiers of computation: methods and new perspectives

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that detector design can be posed as a gradient-descent problem, with first proofs of concept already matching brute-force optimization and new hardware paradigms proposed to scale the method.

desk verdict A useful but thin conference review of differentiable detector design, where the strongest validation claim sits on an unpublished thesis and a duplicated figure caption. read the letter →

arxiv 2501.04448 v1 pith:MKPVV6DV submitted 2025-01-08 hep-ex physics.ins-det

Pietro Vischia This is my paper · ORCID
classification hep-exphysics.ins-det PACS 29.40.-n84.35.+i07.05.Tp
keywords differentiableprogrammingexperimentdesigndetectoroptimizationautomaticdifferentiationneuromorphiccomputingspikingneuralnetworksquantummachinelearninglikelihood-freeinference
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

This paper makes the case that next-generation particle physics experiments can be designed, not just analyzed, with AI: it argues that replacing intractable likelihoods with a differentiable surrogate turns detector design into a gradient-descent problem. It reports first proofs of concept in muon tomography, a parallel-plate avalanche counter for neutron tomography, and a gamma-ray observatory, where optimized design parameters matched or improved on independent brute-force results. It then argues that scaling to collider-sized experiments requires moving beyond conventional CPUs, GPUs, and FPGAs toward spiking neural networks on neuromorphic hardware, and eventually quantum circuits that are analytically differentiable. A sympathetic reader would care because the method promises to explore high-dimensional design spaces that human experts cannot survey, and because it folds cost and geometric constraints directly into the optimization.

What carries the argument

The key object is the design loss in Eq. (1), which integrates an inference and cost penalty over the joint density $p(z|x(\phi),\theta) f(x,\phi)$. Differentiable programming computes exact gradients of this loss via automatic differentiation through the detector simulation, provided the joint density is available in closed form or replaced by a differentiable surrogate such as a neural network. Supporting machinery includes spiking neural networks with leaky integrate-and-fire dynamics for low-power neuromorphic hardware, and quantum circuits whose unitary operations are analytically differentiable, both proposed as routes to scale the gradient loop.

What would settle it

Run the avalanche counter optimization with the paper's method, then evaluate the returned layout with an independent brute-force Monte Carlo scan that does not use the surrogate; if the true performance at the returned parameters falls outside the uncertainty band of the claimed optimum, the central proof of concept is falsified for that case.

Watch

Extended reading notes

Core claim

The paper's central claim is that experimental design can be formulated as an optimization problem whose objective includes both physics performance and cost, and whose gradient can be computed automatically. The author supports this with three proof-of-concept studies: a muon tomography layout, a parallel-plate avalanche counter with optical readout for neutron tomography, and a gamma-ray observatory layout. In the avalanche counter case, the optimized value of one design parameter coincides exactly with the value found by an independent brute-force scan, and the optimization converges to the same solution from different starting points. The paper also argues that conventional hardware cannot scale this gradient loop to LHC-sized experiments, and proposes spiking neural networks on neuromorphic chips and, in the long term, quantum machine learning as the computational route forward.

Load-bearing premise

The entire method assumes that the stochastic relation between design parameters and detector readouts can be written in closed form or replaced by a differentiable surrogate; if that surrogate is biased, the gradient-based optimum will not be the true optimum.

Editorial extensions

If this is right

  • If surrogate-based gradient design is sound, exploring high-dimensional continuous design parameters no longer requires enumerating configurations one by one.
  • The optimizer returns a whole landscape of near-optimal solutions, letting physicists choose a feasible point with domain knowledge instead of blindly accepting a single optimum.
  • The same loss can incorporate arbitrary cost and constraint penalties, so resource limits, geometric constraints, and detector positioning enter the optimization directly.
  • Neuromorphic spiking hardware could make time-pulse-based detector readouts immediately processable at very low power, removing an intermediate reconstruction stage.
  • Quantum machine learning offers analytically differentiable circuits, which in the long term could provide the computational scaling needed for collider-scale experiments.

Reading between the lines

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

  • If the avalanche counter optimum really coincides with a brute-force scan, then in low-dimensional settings the surrogate-based landscape can be trusted; a direct test would be to run the same exact-match check for the muon tomography and gamma observatory cases, which the paper does not report.
  • Because the paper never quantifies surrogate bias, a practical safeguard would be to compare gradients from the surrogate against gradients from full Monte Carlo on a few design points before trusting the optimum.
  • The same machinery could be turned around: a differentiable simulator would allow computing design-parameter sensitivities, not just point optima, helping experimental reviews decide where design precision matters most.
  • A neuromorphic readout would be most convincing if demonstrated in simulation against a full reconstruction chain for the same liquid argon detector, measuring power and latency differences directly.
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

3 major / 4 minor

Summary. This proceedings paper, based on an ICHEP 2024 talk, argues that differentiable programming (automatic differentiation combined with gradient descent) can be used to optimize the design parameters of particle-physics experiments, and that neuromorphic and quantum hardware may make such optimization scalable. The paper gives a short formulation of the design-optimization problem (Eq. (1)), summarizes three proof-of-concept studies (TomOpt for muon tomography, a parallel-plate avalanche counter for neutron tomography, and a gamma-ray observatory layout), and discusses spiking neural networks and quantum machine learning as future paradigms, including a proposal for a neuromorphic readout for the Q-Pix detector.

Significance. The paper is a useful high-level status report on an emerging application of differentiable programming in experimental particle physics. Its main positive points are that it points to the public, peer-reviewed TomOpt code and results, and that it honestly identifies the computational bottleneck (e.g., dedicated CUDA kernels needed even for low-dimensional cases). However, the paper's central quantitative evidence — the claim that an optimized avalanche-counter parameter 'coincides exactly' with an independent brute-force scan — is not verifiable from the manuscript, since the relevant figure caption is erroneous and the supporting document is an unpublished Master's thesis with no public version. The abstract also overstates the neuromorphic part: no neuromorphic implementation is presented, only a proposal. If the hard claims were revised to match the evidence actually shown, the paper would serve as an adequate proceedings contribution; the underlying ideas are promising but the present version does not provide enough detail to independently assess the proofs of concept.

major comments (3)
  1. [Section 5 (Proofs of Concept)] The claim that for the parallel-plate avalanche counter 'the optimization for one of the parameters coincides exactly with the results from independent studies via brute force scan of configurations' is load-bearing for the paper's central assertion that gradient-based design optimization works in practice, but the supporting evidence is not available in the manuscript. Figure 2 is captioned as 'Mean square error of the bias-corrected predictions before and after the optimization loop', identical to Fig. 1's caption, so it does not illustrate the claimed convergence of optimized parameters; Ref. [16] is an unpublished Master's thesis with no public document; and the text does not state whether the brute-force scan uses the same approximate simulator as the gradient pipeline. If both methods share the same simulator, the agreement demonstrates internal consistency rather than physical correctness. Please provide details of the brute-force validation, point to a publicly accessible version of Ref. [16], or temper the exact-coincidence claim accordingly.
  2. [Section 4, Eq. (1)] The optimization problem in Eq. (1) requires the densities p(z|x(φ),θ) and f(x,φ) to be available in closed form or replaced by a differentiable surrogate, but the paper does not quantify the bias of the surrogates used in any of the reported proof-of-concept studies. Without such a quantification, the gradient-based optimum is only known to be optimal with respect to the surrogate, not with respect to the physical simulator. Please add a validation step (e.g., comparing surrogate predictions with high-fidelity Monte Carlo for selected configurations) or explicitly acknowledge this limitation in the discussion of the proofs of concept.
  3. [Abstract and Section 6.1] The abstract states that the paper describes 'first proofs-of-concept of gradient-based optimization of experimental design and implementations in neuromorphic hardware architectures'. However, Section 6.1 contains only a proposal for a neuromorphic readout for the Q-Pix detector, with no implementation, results, or measured performance. The wording should be adjusted to distinguish demonstrated proofs of concept (TomOpt, avalanche counter) from forward-looking proposals (neuromorphic, quantum), so that the abstract accurately reflects the content.
minor comments (4)
  1. [Figure 2] Figure 2's caption is identical to Figure 1's and describes the mean square error before and after the optimization loop, not the convergence of the design parameters; please replace it with a caption that describes what is actually plotted.
  2. [Throughout] There are several typos and grammatical slips, e.g., 'paradigma' in the abstract, 'scarse' in Sec. 1, 'direct acyclic graphs' in Sec. 2, and 'Remarkably converges' in Sec. 5; a careful proofreading pass is needed.
  3. [Section 6.2] The claim that quantum machine learning 'makes it possible to obtain the same accuracy as classical algorithms but using orders of magnitude less data' is a strong statement; as written it lacks the qualifications that are standard in the QML literature (e.g., data-encoding overhead and measurement costs). Please either cite a specific comparative benchmark that supports the claim or soften the wording.
  4. [References] Reference [16] is an unpublished Master's thesis; if it is not publicly available, the reader cannot verify the avalanche-counter results. Consider adding a repository link or citing a published version. Reference [13] is a Zenodo record of a talk; a more archival citation would be preferable for a proceedings paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper summarizes external proof-of-concept studies without reducing claims to fitted inputs or self-citations.

full rationale

This contribution is a short proceedings-style summary of previously published or externally supervised proof-of-concept studies; it contains no derivation whose output equals its input. Equation (1) formalizes the optimization problem and explicitly states the surrogate requirement rather than concealing it, so the framework is not self-definitional. The TomOpt and avalanche-counter results are cited to external works (Refs. [15] and [16]) and are presented as simulations or supervised studies; the avalanche-counter claim of exact coincidence contrasts gradient-based optimization with a brute-force scan, two distinct search procedures, so the agreement is a consistency check of the optimizer rather than a construction-level identity. Self-citations appear (Refs. [3], [13], [16]) but they are contextual and not load-bearing: the prior-art point about machine-learned likelihood ratios is supported by independent BaBar and CMS references, and the proof-of-concept claims rest on external simulations rather than on the author's own assertions. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is imported through self-citation. The lack of quantified surrogate bias is a validation and correctness gap, not a circularity pattern, because the paper does not claim that the surrogate and the target are identical by construction. Therefore no circular steps are identified.

Assumptions & free parameters 0 free parameters · 3 assumptions · 1 invented entities

The paper performs no fitting or parameter estimation; it reviews existing studies. It relies on the domain assumption that differentiable surrogates accurately represent the stochastic simulation, and on the ad hoc assumption that the cited examples are representative. One speculative entity, the neuromorphic Q-Pix readout, is introduced without evidence.

assumptions (3)
  • domain assumption The stochastic simulation pipeline can be expressed as a differentiable function, either in closed form or via an accurate surrogate model.
    Equation (1) requires p(z|x(phi),theta) f(x,phi) to be writable in closed form or replaced by a differentiable surrogate; if this fails, gradient descent cannot be applied.
  • domain assumption Gradient descent on the surrogate loss yields the true optimum of the physical design objective.
    The paper assumes that optimizing the differentiable approximation gives designs that optimize the actual intractable likelihood; no proof or error bound is given.
  • ad hoc to paper The reported proofs of concept are representative and not cherry-picked.
    The paper selects a few examples from its own group's work (TomOpt, avalanche counter, gamma ray observatory) without a systematic survey, so the 'first proofs of concept' claim may be incomplete.
invented entities (1)
  • Neuromorphic readout for Q-Pix
    purpose: Make Q-Pix time-dependent pulses immediately processable without intermediate encoding or reconstruction.
    Proposed in Section 6.1 with no implementation, simulation, or measurement; no falsifiable handle is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AI-assisted design of experiments at the frontiers of computation: methods and new perspectives." pith.science (2026). https://pith.science/paper/MKPVV6DV

@misc{pith2026250104448,
  author       = {Pith},
  title        = {Pith review of: AI-assisted design of experiments at the frontiers of computation: methods and new perspectives},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MKPVV6DV}},
  note         = {Machine review of arXiv:2501.04448}
}
read the original abstract

Designing the next generation colliders and detectors involves solving optimization problems in high-dimensional spaces where the optimal solutions may nest in regions that even a team of expert humans would not explore. Resorting to Artificial Intelligence to assist the experimental design introduces however significant computational challenges in terms of generation and processing of the data required to perform such optimizations: from the software point of view, differentiable programming makes the exploration of such spaces with gradient descent feasible; from the hardware point of view, the complexity of the resulting models and their optimization is prohibitive. To scale up to the complexity of the typical HEP collider experiment, a change in paradigma is required. In this contribution I will describe the first proofs-of-concept of gradient-based optimization of experimental design and implementations in neuromorphic hardware architectures, paving the way to more complex challenges.

Figures

Figures reproduced from arXiv: 2501.04448 by the authors.

Figure 1
Figure 1. Mean square error of the bias-corrected predictions before and after the optimization loop. Reproduced with authorization from Ref. [15]. of solutions around the minimum of the loss function: the physicist then will use her judgment and domain knowledge to choose the most feasible solution, integrating in the decision any consideration that was not parameterized in the model, either because of discreteness (discrete… view at source ↗
Figure 2
Figure 2. Mean square error of the bias-corrected predictions before and after the optimization loop. Reproduced with authorization from Ref. [16]. 6. New computational paradigms The proofs of concept described in Section 5 are computational challenging even in these low-dimensional cases. For instance, computational tricks (such as dedicated CUDA kernels) had to be devised to make the optimization in TomOpt feasible for suit… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

  1. [16]

    Automatic Optimization of a Parallel-Plate Avalanche Counter with Optical Readout

    Pereira Martínez, M., Cid Vidal, X. (sup.), Vischia, P. (sup.), "Automatic Optimization of a Parallel-Plate Avalanche Counter with Optical Readout". Master Thesis defended in 2024 at Universidad de Santiago de Compostela

    work page 2024

  2. [1]

    2020 Update of the European Strategy for Particle Physics

    The European Strategy Group, "2020 Update of the European Strategy for Particle Physics" (Brochure). (2020), https://cds.cern.ch/record/2721370

    work page arXiv 2020

  3. [2]

    Reporting results in High Energy Physics publications: A manifesto

    Vischia, P., "Reporting results in High Energy Physics publications: A manifesto".Reviews In Physics.5 pp. 100046 (2020)

    work page 2020

  4. [3]

    Toward the end-to-end optimization of particle physics instruments with differentiable programming

    Dorigo, T., Giammanco, A., Vischia, P. (eds.), et al. , "Toward the end-to-end optimization of particle physics instruments with differentiable programming".Reviews In Physics. 10 pp. 100085 (2023)

    work page 2023

  5. [4]

    The Nature of Statistical Learning Theory

    Vapnik, V., "The Nature of Statistical Learning Theory". (Springer Nature, 1999)

    work page 1999

  6. [5]

    Mathematics of Deep Learning

    Vidal, R., et al., "Mathematics of Deep Learning". (2017), https://arxiv.org/abs/1712.04741

    work page Pith review arXiv 2017

  7. [6]

    Evaluating Derivatives

    Griewank, A. & Walther, A., "Evaluating Derivatives". (Society for Industrial,2008)

    work page 2008

  8. [7]

    A Logical Calculus of Ideas Immanent in Nervous Activity

    Mcculloch, W. & Pitts, W., "A Logical Calculus of Ideas Immanent in Nervous Activity." Bulletin Of Mathematical Biophysics. 5pp. 127-147 (1943)

    work page 1943

Show all 24 references
  1. [8]

    DeepLearningestmort

    LeCun,Y.,"DeepLearningestmort".(https://www.facebook.com/yann.lecun/posts/10155003011462143), Accessed: 2023-05-23

    2023

  2. [9]

    Making the world differentiable: on using self supervised fully recurrent neural networks for dynamic reinforcement learning and planning in non-stationary environ- ments

    Schmidhuber, J., "Making the world differentiable: on using self supervised fully recurrent neural networks for dynamic reinforcement learning and planning in non-stationary environ- ments". (Inst. für Informatik,1990) 6 AI-assisted design of experiments at the frontiers of co...

    1990

  3. [10]

    Amethodforapproximatingoptimalstatisticalsignificanceswithmachine- learned likelihoods

    Arganda,E.,etal.,"Amethodforapproximatingoptimalstatisticalsignificanceswithmachine- learned likelihoods". Eur. Phys. J. C 82, 993 (2022)

    2022

  4. [11]

    MeasurementofCPasymmetriesandbranchingfractionsincharm- less two-body B-meson decays to pions and kaons

    TheBaBarCollaboration,"MeasurementofCPasymmetriesandbranchingfractionsincharm- less two-body B-meson decays to pions and kaons".Phys. Rev. D.87, 052009 (2013,3)

    2013

  5. [12]

    Search for 𝑡𝑡𝐻 production in the 𝐻 → 𝑏𝑏 decay chan- nel with √𝑠 = 13 TeV pp collisions at the CMS experiment

    The CMS Collaboration, "Search for 𝑡𝑡𝐻 production in the 𝐻 → 𝑏𝑏 decay chan- nel with √𝑠 = 13 TeV pp collisions at the CMS experiment". (CERN,2016), http://cds.cern.ch/record/2139578

    2016

  6. [13]

    Machine Learning and Differentiable Programming for Statistics at the LHC

    Vischia, P., "Machine Learning and Differentiable Programming for Statistics at the LHC". (https://doi.org/10.5281/zenodo.13963379 and https://indico.cern.ch/event/1198609/contributions/5366510/)

  7. [14]

    UnveilingtheStructureofWideFlatMinimainNeuralNetworks

    Baldassi, C., etal., "UnveilingtheStructureofWideFlatMinimainNeuralNetworks". Phys. Rev. Lett.. 127, 278301 (2021,12)

    2021

  8. [15]

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

    Strong, G., et al., "TomOpt: differential optimisation for task- and constraint-aware design of particle detectors in the context of muon tomography".Machine Learning: Science And Technology.5, 035002 (2024,7)

    2024

  9. [17]

    End-To-End Optimization of the Layout of a Gamma Ray Observatory

    Dorigo, T., Aehle, M., Donini, J., Doro, M., Gauger, N., Izbicki, R., Lee, A., Masserano, L., Nardi, F., S, S. & Shen, A., "End-To-End Optimization of the Layout of a Gamma Ray Observatory". (2023)

    2023

  10. [18]

    Modelling the neurons of the electrosensory lobe in Gym- notus omarorum with differentiable programming

    Vischia, P., Caputi, A. A. (2023), "Modelling the neurons of the electrosensory lobe in Gym- notus omarorum with differentiable programming". Neurocoding 2023, Piriápolis, Uruguay. https://doi.org/10.5281/zenodo.8394819

    2023 doi

  11. [19]

    Nat Commun 15, 3392 (2024)

    Pal,A.,Chai,Z.,Jiang,J.etal.Anultraenergy-efficienthardwareplatformforneuromorphic computing enabled by 2D-TMD tunnel-FETs. Nat Commun 15, 3392 (2024)

    2024

  12. [20]

    Neuromorphic Readout For Homogeneous Hadron Calorimeters

    Lupi, E., et al., "Neuromorphic Readout For Homogeneous Hadron Calorimeters". (https://indico.cern.ch/event/1380163/contributions/6102429/)

  13. [21]

    Enhanced low-energy supernova burst detection in large liquid argon time projection chambers enabled by Q-Pix

    Kubota, S., et al., "Enhanced low-energy supernova burst detection in large liquid argon time projection chambers enabled by Q-Pix".Phys. Rev. D.106, 032011 (2022,8)

    2022

  14. [22]

    MachineLearningwithQuantumComputers

    Schuld,M.,Petruccione,F.,"MachineLearningwithQuantumComputers".(SpringerNature, 2021)

    2021

  15. [23]

    Event Classification with Quantum Machine Learning in High-Energy Physics

    Terashi, K., Kaneda, M., Kishimoto, T. et al., "Event Classification with Quantum Machine Learning in High-Energy Physics". Comput Softw Big Sci 5, 2 (2021). 7 AI-assisted design of experiments at the frontiers of computation: methods and new perspectives Pietro Vischia

    2021

  16. [24]

    Sentiment Analysis with QuantumNaturalLanguageProcessing

    Uría García, M., Vischia, P. (sup.), Uría González, J.M. (sup.), "Sentiment Analysis with QuantumNaturalLanguageProcessing".BachelorThesisdefendedin2024atUniversidadde Oviedo. 8

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.