A Differentiable Surrogate for Loss-Dominated Ion Beam Transport

arXiv:2610.08877 · physics.ins-det, nucl-ex · Submitted 2026-10-06 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "A Differentiable Surrogate for Loss-Dominated Ion Beam Transport".

Paper summary: Kai: So, looking at the whole picture of this paper, "A Differentiable Surrogate for Loss-Dominated Ion Beam Transport," the authors are proposing a method to model ion beam transport when particle loss is a big factor.

Mira: They claim this surrogate jointly captures the survival probability and phase space distribution by using discrete-time survival analysis and conditional normalizing flows to create one unified likelihood function for observations.

Lev: The implication for users is that they can use this model to optimize operating voltages through gradient ascent, which bypasses the need for running order one hundred eight trajectories for every configuration in a black-box optimization <ref:2610.08877#pg1>.

Kai: It means you can design or tune your beamline much faster because you have a fast approximation to the map from operating conditions to physical observables.

Mira: The paper shows that when trained on simulated data, this surrogate provides an inference-ready model for experimental design and online tuning in accelerator physics.

Lev: The authors also mention that they are extending this approach as a digital twin testbed at MIT and Texas A andM University for nuclear precision experiments.

Kai: It seems like the core idea is taking these hard loss scenarios where simulations are slow or non-differentiable and turning them into something you can actually use to find better settings in the real world.

Conclusion: Kai: So we're wrapping up on "A Differentiable Surrogate for Loss-Dominated Ion Beam Transport." Basically, they've built a way to make these particle tracking simulations differentiable so you can optimize the voltages directly through gradient ascent.

Mira: That’s the core idea, right? It takes simulations where particles are lost and turns them into a probabilistic model that gives us gradients. The authors use discrete time survival analysis and conditional normalizing flows to get this unified likelihood function for both particle counts and phase space.

Lev: From an engineering side, what this means is you can tune the operating voltages using standard optimization routines without needing to run hundreds of full simulations for every small change in the settings. It makes inference much faster.

Kai: Right, so it moves particle tracking from a slow simulation step to something that can guide experimental design in real time. The results show they get high accuracy on both survival and phase space emulation when trained on about sixty percent of the data.

Mira: They also showed that this method can recover the surviving mass with an R squared value above zero point ninety-four at every detector plane, and it tracks the shot noise floor error along the beamline pretty well.

Lev: The caveat is that they noted a specific mismatch when you condition on complete six dimensional phase space clouds. Even then, because of that residual emulator to SIMION mismatch, the voltage posterior isn't perfectly centered at that statistical power level.

Kai: So it’s a very useful tool for getting initial guesses, but you still need to account for the simulation error in the final precision mapping. It’s a powerful inference model, but we’re still chasing that perfect link between simulation and experiment.

Mira: Exactly, and that leads us into what comes next: how do we actually make this surrogate work when we start looking at real experimental data, not just simulated SIMION data?

J. M. Munoz, L. Ruiz-Arango, D. Gonzalez-Acevedo, F. C. Pastrana-Cruz, J. Hacias, R. F., Garcia Ruiz

Laboratory for Nuclear Science, Massachusetts Institute of Technology

physics.ins-det, nucl-ex

Submitted: 2026-10-06

Updated: 2026-11-15

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 90/100

The gist: A differentiable probabilistic surrogate for ion beam transport in regimes dominated by particle loss introduces a model that combines discrete-time survival analysis with conditional normalizing…

Key concepts

Discrete-time survival analysis
This technique is used to model particle loss over time or across different planes in the beamline. It treats each plane as a step, calculating the probability that an ion survives from one plane to the next based on a learnable 'hazard function.' This helps quantify how many ions are expected to survive.
Conditional normalizing flows
Normalizing flows are used to model the complex six-dimensional phase-space distribution of surviving ions in any given plane. They take a simple, known distribution (like a Gaussian) and transform it into the actual, more complicated phase-space data observed at that plane, conditioned on factors like electrode voltages.
Differentiable surrogate
The model is designed to be differentiable so that standard gradient-based optimization methods can be used. This means researchers can directly optimize operating parameters (like voltage settings) by calculating how changes in those parameters affect the final predicted outcome, even when the underlying simulation involves hard losses.
Likelihood function
This is a mathematical function that calculates the probability of observing the actual detector counts and phase-space measurements given a specific set of model parameters (like voltages). The paper creates a tractable likelihood by linking particle survival probabilities and flow marginals, making it usable for inference.

Terminology

Summary

A differentiable probabilistic surrogate for ion beam transport in regimes dominated by particle loss introduces a model that combines discrete-time survival analysis with conditional normalizing flows to enable gradient-based optimization and inference from sparse detector observations

Model Introduction and Approach

The paper introduces a differentiable probabilistic surrogate for ion beam transport in regimes dominated by particle loss This model factorizes the mixed simulator response into the probability of survival to each detector plane and the conditional phase-space distribution of surviving ions The approach combines discrete-time survival analysis with a conditional normalizing flow to yield a unified, tractable likelihood for particle counts and phase-space observations

Data Generation and Model Heads

The particle-tracking dataset is generated using SIMION simulations for each electrode-voltage configuration, recording both the number of ions that survive to that location and the six-dimensional phase-space coordinates of the surviving particles The ion transport is modeled with a multiplicative ansatz over the surviving mass and beam shape, defined by Equation (1) The transmission head models particle loss using discrete-time survival analysis where the learnable hazard function is modeled as the probability that an ion admitted to plane k − 1 is lost before plane k The phase-space head models the kinetic ion distribution in a given plane as a normalizing flow, which carries a Gaussian base variable to the phase-space at any plane, conditioned on the electrode voltages and arc length

Optimization and Inference Capabilities

The surrogate enables gradient-based optimization of operating voltages by making the transmission head differentiable in the voltages, allowing log T10 to be ascended directly through a bagged ensemble This process transforms particle-tracking simulations with hard losses into an inference-ready model for experimental design and beamline optimization Furthermore, the likelihood enables inference of the beam in planes without a detector by transforming Eq. (1) into a tractable likelihood function involving binomials and flow marginals

Performance Benchmarks

The model demonstrates high accuracy in both ion-survival and phase-space emulation when trained on a 60% subset of the simulated configurations The transmission head recovers the surviving mass with an R2 staying above 0.94 at every detector, and its error tracks the binomial shot-noise floor along the beamline The phase-space head reproduces SIMION phase space spread, showing that at det6, immediately after the strongest cut in the geometry, emittance grows relative to SIMION and the centroid is offset by a third of the SIMION spread The constrained optimization variant successfully finds settings that are feasible within 5% of the target transmission when optimizing under beam-width constraints The inference procedure shows that conditioning on complete six-dimensional phase-space clouds tightens the voltage posterior further but leaves it mis-centred, because at that statistical power the residual emulator–SIMION mismatch dominates the inference The energy-distance z-score for the flow is lower than for other models, indicating better performance in reconstructing kinetic-energy marginals

Future Directions and Challenges

The main remaining challenge is to close the gap between simulation and experiment by linking the survival and beam-shape models more closely Future work should focus on calibrating the surrogate with experimental data to account for imperfect field maps, electrode misalignments, and environmental drift This approach is being extended to nuclear precision experiments at MIT and Texas A&M University as testbeds for a digital twin The paper concludes that this differentiable probabilistic surrogate provides millisecond-scale differentiable predictions, transforming particle-tracking simulations with hard losses into an inference-ready model

Acknowledgments

The authors acknowledge support from the PCLB Foundation and the U.S. Department of Energy, Office of Science, under Award Number DE-SC0026653 (Genesis Mission) Computational resources were provided by SubMIT at MIT Physics The authors thank Scott Moroch, Alex Brinston, and Arian Jadbabaie (MIT), Jonas Karthein (Texas A&M University), and the DOE Genesis FastSims project for insightful discussions

References

[1] Silviu-Marian Udrescu, Antoine Belley, Jason D. Holt, Gilad Perez, and Ronald F. p. 5 Garcia Ruiz. p. 0 Probing the electroweak structure of nuclei with rare atoms and molecules

[2] X F Yang, S J Wang, S G Wilkins, and R F Garcia Ruiz. Laser spectroscopy for the study of exotic nuclei

[3] Gordon Arrowsmith-Kron et al. Opportunities for fundamental physics research with radioactive molecules

[4] Helmut Wiedemann. Particle Accelerator Physics

[5] Wayne D. Newhauser and Rui Zhang. p. 5 The physics of proton therapy

[6] Francis F. Chen. Introduction to Plasma Physics and Controlled Fusion

[7] T. Kalvas, O. Tarvainen, T. p. 5 Ropponen, O. p. 5 Steczkiewicz, J. p. 5 Ärje, and H. p. 5 Clark. IBSIMU: A threedimensional simulation software for charged particle optics

[8] Andreas Adelmann, Pedro Calvo, Matthias Frey, Achim Gsell, Uldis Locans, Christof Metzger-Kraus, Nicole Neveu, Chris Rogers, Steve Russell, Suzanne Sheehy, Jochem Snuverink, and Daniel Winklehner. OPAL a versatile tool for charged particle accelerator simulations

[9] Allison McCarn Deiana et al. Applications and techniques for fast machine learning in science

[10] Anastasia Pipi, Valentin Duruisseaux, Emily Been, Xuecheng Tao, Taylor L. Patti, Anima Anandkumar, and Prineha Narang. p. 5 Inverse design of quantum control sequences with fourier neural operators

[11] Jose M. Munoz et al. p. 0 Linking electromagnetic moments to nuclear interactions with a global physics-driven machine-learning emulator

[12] Elena Cuoco, Marco Cavaglià, Ik Siong Heng, David Keitel, and Christopher Messenger. Applications of machine learning in gravitational-wave research with current interferometric detectors: E. p. 6 cuoco et al. p. 6 Living Reviews in Relativity

[13] Auralee Edelen et al. Opportunities in machine learning for particle accelerators

[14] Auralee Edelen, Nicole Neveu, Matthias Frey, Yannick Huber, Christopher Mayes, and Andreas Adelmann. Machine learning for orders of magnitude speedup in multiobjective optimization of particle accelerator systems

[15] J. Duris et al. p. 6 Bayesian optimization of a free-electron laser

[16] Ryan Roussel, Auralee Edelen, Christopher Mayes, Daniel Ratner, Juan Pablo Gonzalez-Aguilera, Seongyeol Kim, Eric Wisniewski, and John Power. Phase space reconstruction from accelerator beam measurements using neural networks and differentiable simulations

[17] Alexander Scheinker. Conditional guided generative diffusion for particle accelerator beam diagnostics

[18] Michela Paganini, Luke de Oliveira, and Benjamin Nachman. CaloGAN: Simulating 3D high energy particle showers in multilayer electromagnetic calorimeters with generative adversarial networks

[19] Baran Hashemi and Claudius Krause. Deep generative models for detector signature simulation: A taxonomic review

[20] Ross L. Prentice and Lynn A. p. 6 Gloeckler. Regression analysis of grouped survival data with application to breast cancer data

[21] Gerhard Tutz and Matthias Schmid. Modeling Discrete Time-to-Event Data

[22] Conor Durkan, Artur Bekasov, Iain Murray, and George Papamakarios. Neural spline flows

[23] David A. Dahl. SIMION for the personal computer in reflection

[24] Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder–decoder for statistical machine translation

[25] Michael F. Gensheimer and Balasubramanian Narasimhan. p. 6 A scalable discrete-time survival model for neural networks

[26] George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. Normalizing flows for probabilistic modeling and inference

[27] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using Real NVP

[28] Matthew Tancik et al. Fourier features let networks learn high frequency functions in low dimensional domains

[29] Diederik P. Kingma and Jimmy Ba. p. 6 Adam: A method for stochastic optimization

[30] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution

[31] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models

[32] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models

[33] George Papamakarios, David C. Sterratt, and Iain Murray. p. 7 Sequential neural likelihood: Fast likelihood-free inference with autoregressive flows

[34] M. S. Albergo, G. p. 7 Kanwar, and P. p. 7 E. Shanahan. Flow-based generative models for Markov chain Monte Carlo in lattice field theory

[35] Frank Noé, Simon Olsson, Jonas Köhler, and Hao Wu.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Surrogate-based constrained optimization allows for gradient ascent of operating conditions by predicting log T10 is ascended directly through a bagged ensemble which can drive voltages to settings that exceed previous optimization results, as seen in The predicted maximum from gradient ascent achieved up to 19.2% when validated in SIMION, above the 16.7% best setting found during the black-box optimization campaign used in data generation.

  2. Inference of unobserved beam properties is enabled by a likelihood function that reconstructs phase space at planes without detectors using Eq. (3) into p(D u) = Ym i=1 Bin nki; nki−1, Tki Tki−1 MYki j=1 pθ y(ki) j u, ski. This enables the reconstruction of the ion beam in planes without a detector, conditioned on observed counts and exit images.

  3. The system provides millisecond-scale differentiable predictions that transform particle-tracking simulations with hard losses into an inference-ready model for experimental design, beamline optimization, and future digital twins of experiments with rare nuclei. This capability allows for rapid evaluation of operating conditions compared to traditional methods which require order 108 ion trajectories.

  4. The phase-space head can impose constraints beyond scalar transmission by demonstrating this through imposing an augmented-Lagrangian equality on the predicted beam width at det5, which SIMION validation shows the delivered width decreases monotonically with the target. This enables optimization under beam-shape constraints.

Sources

Related papers