Reinforcement Learning in High-Energy Physics: A Survey from Accelerator Control to Theory Exploration
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.
Kai: We're looking at the paper: "Reinforcement Learning in High-Energy Physics".
Mira: Here's the gist of the paper. Reinforcement learning (RL) is a machine learning paradigm where an agent learns a decision-making policy through repeated interaction with an environment, guided by a scalar reward signal rather than explicit supervision.
Kai: This approach is particularly advantageous in High Energy Physics (HEP) tasks that are inherently closed-loop decisions, involve sparse or delayed feedback, and require objectives measured on real machines or stochastic simulations—situations where traditional supervised or unsupervised learning methods are less suitable.
Mira: RL excels in scenarios involving repeated tasks with fast simulators, decisions within fast control loops, operation under faults or corrupted measurements, the use of physics-based rewards instead of labels (such as scattering likelihoods), and sequential discrete constructions with a reusable policy.
Kai: This survey provides a comprehensive review of RL applications across seven major domains in HEP: 1.
Title and authors: Kai: So we're talking about this paper, "Reinforcement Learning in High-Energy Physics: A Survey from Accelerator Control to Theory Exploration". It covers a lot of ground, starting right from how agents learn decisions through interaction with an environment guided by a reward signal instead of just having some pre-written supervision.
Mira: Exactly. It sets up this paradigm where the data the algorithm observes depends on its choices, and that feedback can be sparse or delayed when you're dealing with things like real machines or complex simulations.
Lev: And for us in the quantum hardware world, that means we're looking at how this works when we want to deploy agents in real-time feedback loops for accelerator control, which could mean microsecond inference latency on dedicated hardware like FPGAs to reject disturbances and maintain stable regulation under time-varying currents.
Kai: Right, so it’s not just theory; it’s about what we can actually build and measure. The survey looks at applications across seven major domains in HEP.
Mira: It touches everything from accelerator control and beam optimization to data quality monitoring, jet physics, charged particle tracking, beyond the standard model exploration, detector design, and even formal quantum field theory computation.
Lev: That breadth is what’s interesting—it shows how RL isn't just a niche tool for one problem but something that spans the whole experimental process. I wonder how we actually get those physics-based rewards defined when you're dealing with such complex systems, Kai?
Kai: That’s the first big hurdle. The paper emphasizes that embedding domain knowledge into state representation and reward design is absolutely critical, because without it, you just get something called reward hacking.
Mira: It really does. We need to move beyond just plugging in a formula for a reward and actually figure out what a good decision looks like in physics terms before the agent even starts learning.
Lev: From my side, if we try to run these on real hardware, we’re looking at sample efficiency being a huge factor because getting enough interactions to train the policy can be tough when you have complex environments.
Title and authors: Kai: The paper also discusses various RL methodologies, not just the standard model-free methods like Deep Q networks or policy gradient methods, but also model-based planning techniques like Dyna and Monte Carlo tree search.
Mira: And it brings up things like multi-agent RL, inverse learning where you try to imitate an expert’s behavior, and even bandit algorithms for tasks where you don't have labels at all.
Lev: When you look at sample efficiency, that ties directly into deployment; we need strategies like replay buffers that keep the agent learning from past experiences without needing constant real-time interaction with the physical system.
Kai: The survey also stresses practical considerations like sim-to-real transfer techniques, such as domain randomization to bridge those gaps between simulation and reality where you can't just build a perfect digital twin.
Mira: And safety is huge too; they mention Constrained MDPs, or CMDPs, which are ways to explicitly model operational constraints so the agent doesn't just learn something that violates physical limits.
Lev: Those safety mechanisms are essential for anything that touches real hardware; you can’t afford an agent making a decision that compromises beam stability or detector integrity because of some reward glitch.
Kai: The paper highlights the common first obstacle: formalizing a physics task as a proper MDP, shifting from static learning to these dynamic decision processes guided by those reward functions.
Mira: And what I find important is their discussion on how RL has shown promise in matching or even improving upon expert designs across areas like jet grooming and track finding.
Lev: I agree on the matching part. When we look at jet grooming, for instance, RL can learn optimal node-by-node decisions along a clustering tree based on physics rewards, which improves the downstream classifier performance significantly.
Kai: That’s what they mean by using RL for jet physics—learning those specific decisions that matter for the final analysis. But it also covers things like data caching and real-time trigger control, which are crucial for managing background composition changes during a fill.
Title and authors: Mira: It’s interesting to see it reach into formal quantum field theory computation as well, using graph and grammar based RL to explore discrete spaces of consistent theories like string compactifications.
Lev: That moves the concept from just controlling detectors to exploring theoretical models that might be too complex for traditional search methods. It suggests a new way to navigate those discrete spaces.
Kai: So, looking at the conclusion of this paper, they summarize what RL has achieved—it’s shown success in many areas—but they clearly outline the open challenges going forward.
Mira: They point out that robust reward design is still a major issue and we still lack shared benchmarks for cross-section measurements where you can really test these policies.
Lev: And sim-to-real transfer remains a big sticking point; we need better ways to ensure what works in simulation actually translates reliably to the noisy, physical world.
Kai: The paper ends by stressing that while RL has shown promise, it’s not a universal replacement for everything; we still need formal safety guarantees and statistical validity checks for decisions made by these policies.
Mira: It’s a strong overview of where the field is right now—it shows the potential but also lays out exactly what needs to be done next to make RL reliable in HEP.
Lev: For me, I think the real path forward involves integrating physics knowledge more deeply into the state representation itself, moving away from purely data-driven learning toward physics-informed RL.
Kai: That sounds like a solid direction for future work—making sure the agent understands *why* it's making a certain choice based on physical principles, not just statistical correlation.
Mira: Exactly. The next step is less about finding the best algorithm and more about creating the best reward function and state definition for the specific physics problem you’re trying to solve.
Lev: And that brings us nicely into what I think we should look at next, because if we can figure out those simulation gaps, it opens up a lot of possibilities for real-world deployment.
The paper's summary: Kai: So, we just finished looking at this survey paper on Reinforcement Learning in High-Energy Physics. It really paints a picture of where the field is right now—it’s not just one thing anymore, it’s everywhere.
Mira: Exactly. The authors show how RL isn't just some neat trick for one problem; it spans seven huge areas in HEP, from controlling particle accelerators to figuring out how to build new theories.
Lev: I mean, the survey lays out a lot of methodologies they look at—model free stuff like Deep Q networks, policy gradient methods like PPO and REINFORCE, and even model based planning things like Dyna.
Kai: Yeah, but what's really interesting is that it hammers home the same core challenges across all those domains. The first big hurdle is formalizing a physics task as an Markov Decision Process. It’s moving from static learning to these dynamic decision processes guided by reward functions.
Mira: And the authors stress that you absolutely have to embed domain knowledge into the state representation, the reward design, and the action space constraints, otherwise you get reward hacking. That’s a major warning for anyone trying to apply this in practice.
Lev: From my side, it reminds me of how hard it is to get that physics-informed reward right when you're dealing with real hardware; if the agent learns something statistically good but physically impossible, the whole setup fails.
Kai: And they talk a lot about practical things like sim-to-real transfer using domain randomization to bridge the gap between simulation and reality. That’s where most of the trouble happens when you move from a clean computer to a messy detector.
Mira: They also mention safety mechanisms, specifically Constrained Markov Decision Processes, or CMDPs. That’s about making sure the agent doesn't learn something that violates operational constraints, like pushing the beam too hard.
Lev: Which brings us to sample efficiency—getting enough training data is tough when you have complex environments, so they talk about replay buffers and other strategies to keep the learning going without constantly hitting the real thing.
Kai: It’s a lot of moving parts then—from basic accelerator steering to optimizing jet physics and even exploring Beyond the Standard Model theories using graph-based RL. It shows a lot of promise in matching or improving upon expert designs where those complex decisions are involved.
Mira: But they aren't saying it’s a magic solution for everything. The paper concludes by pointing out the big open challenges: we still need better reward design, more robust sim-to-real transfer techniques, and frankly, shared benchmarks so we can really test these policies statistically.
Lev: So it sounds like the consensus is that RL is a powerful tool for decision making in HEP, but it’s not a universal replacement for traditional methods yet. We need to focus on those safety guarantees and making sure the statistical validity of its decisions holds up for things like cross-section measurements.
Kai: It’s definitely a big conversation because if we can solve those simulation gaps and get better reward design, this could fundamentally change how we approach complex experimental data analysis over time.
The paper's improvements: Tom: So we’re looking at how the authors suggest we move forward with this Reinforcement Learning survey paper. They aren't just stopping at what RL can do now; they're giving us a roadmap for what needs to happen next across all these seven HEP areas.
Kai: Right, they emphasize that the common first obstacle is formalizing physics tasks as MDPs, and the authors suggest we need to get much better at that state representation and reward design. It’s about moving past just plugging in a formula for a reward and actually making the agent understand the physics behind its choices.
Mira: I agree with Kai there because if you don't have good domain knowledge baked into those inputs, the whole learning process breaks down. The survey really pushes that idea that we need techniques like pretraining policies before fine-tuning them for specific tasks.
Lev: From an engineering standpoint, I see their focus on sim-to-real transfer—domain randomization—as crucial because you can't just build a perfect digital twin of a complex detector or accelerator system. You have to learn how to handle the messy physics in reality.
Kai: And they talk about safety again, suggesting we need better Constrained MDPs so that the agents don't learn something physically impossible or dangerous when operating real machines. It’s not just about getting a good result; it’s about keeping things stable.
Mira: That links back to my point on assumptions. The survey suggests that for BSM model building, using graph and grammar based RL is a way to navigate those discrete spaces of symmetry groups and compactification data more effectively than traditional search methods.
Lev: That sounds promising for theory exploration because it lets you systematically explore configurations that might be too complicated for standard searches. It moves the exploration from guessing to learning a path through the space itself.
Kai: So, overall, the improvements they suggest boil down to three main things: better physics knowledge integration, more robust sim-to-real transfer methods, and stronger safety constraints built directly into the learning process.
Mira: I think those are exactly what's needed for real deployment; you need policies that are not just mathematically sound but physically plausible and safe in a real experimental environment.
Lev: It also points toward the future needing shared benchmarks so we can actually compare how well these RL policies perform against established methods. We need to know if the agent is actually doing better than a human expert in those jet grooming or track finding tasks they mentioned earlier.
Kai: So, it seems like the next big step isn't just more algorithms, but building better tools that bridge the gap between abstract theory and reliable real-world operation.
Conclusion: Kai: So we’re wrapping up this discussion on "Reinforcement Learning in High-Energy Physics: A Survey from Accelerator Control to Theory Exploration". It boils down to this: RL is a powerful decision tool, but its real value depends entirely on how well we design the physics problem it’s trying to solve.
Mira: Right. The paper shows that we have these different methods—model free, policy gradient, and model based planning—but the authors stress that the real hurdle isn't just choosing the right algorithm; it’s getting the reward function exactly right.
Lev: I think what this survey really tells us is that moving toward practical deployment means we have to prioritize sim-to-real transfer techniques like domain randomization because you can't just build a perfect digital world for these agents.
Kai: It’s true, and I think the engineers need to pay close attention to those safety constraints they discuss, like Constrained MDPs, because when you’re controlling real hardware, stability is more important than getting a high score in simulation.
Mira: And for the theorists looking at BSM models, it suggests that graph and grammar based RL could be a way to explore those discrete spaces of theories in a systematic way that traditional searching just can't handle.
Lev: If we look at the numbers, getting enough data for these policies to learn reliably is still tough because you’re dealing with complex environments where you need massive sample counts.
Kai: So, the main implication for us is that RL isn't a magic fix; it’s a sophisticated decision-making layer that requires deep physics knowledge woven into every step of the process.
Mira: Precisely. It changes how we think about these tasks—not as static calculations, but as dynamic processes where the agent learns to make sequential decisions under uncertainty.
Lev: For building real hardware, it means we need to focus our error correction efforts on making sure the policies don't introduce instabilities when they interact with physical systems.
Kai: So this survey of "Reinforcement Learning in High-Energy Physics: A Survey from Accelerator Control to Theory Exploration" gives us a really solid overview of the potential across all these experimental domains.
Mira: It’s a great map, but it also clearly shows where the research needs to go next, especially regarding shared benchmarks and more rigorous testing of those reward designs.
Lev: For me, it highlights that if we can get better at modeling those complex systems through RL, it opens up new ways to tackle things like real-time trigger control and data caching.
Kai: Exactly. It’s a lot of foundational work, but it really sets the stage for what we can build next in the quantum hardware and experimental side.
Mira: So yeah, that covers the main points of this survey on Reinforcement Learning in High-Energy Physics. We’ll be keeping an eye on those open challenges as they get addressed.
Luca Anzalone
hep-ex, cs.LG, hep-ph
Submitted: 2026-11-01
Updated: 2026-11-01
Comments: 86 pages, 5 figures, 3 tables
Code: https://github.com/SebastianMacaluso/ToyJetsShower
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 86/100
The gist: Reinforcement learning (RL) is a machine learning paradigm where an agent learns a decision-making policy through repeated interaction with an environment…
Key concepts
- MDP Formulation
- The core challenge is turning a physical task into a mathematical decision process called a Markov Decision Process (MDP). This means defining states (the current situation), actions (what the agent can do), and rewards (how good or bad that action was). This step shifts learning from simple pattern recognition to sequential decision-making guided by physics objectives.
- Reward Design
- Since RL learns from rewards, designing them correctly is critical. The paper stresses avoiding 'reward hacking'—where the agent finds unintended loopholes to maximize the score without achieving the actual physical goal. Rewards must be carefully crafted using physics knowledge to guide agents toward desired outcomes.
- Sim-to-Real Transfer
- RL often trains in fast simulations, but deployment requires real machines. Sim-to-real transfer techniques, like domain randomization, are necessary to bridge the gap between simulated environments and physical reality. This ensures that a policy learned in simulation performs reliably when interacting with actual experimental hardware.
- Physics-Informed RL (PIRL)
- This approach integrates known physics principles directly into the RL framework. Instead of relying solely on trial and error, agents are guided by constraints or rewards derived from established physical laws. This embedding of domain knowledge helps ensure that the learned policies make physically plausible decisions.
Terminology
Summary
Reinforcement learning (RL) is a machine learning paradigm where an agent learns a decision-making policy through repeated interaction with an environment, guided by a scalar reward signal rather than explicit supervision. p. 0 anzalone@proton. me
Abstract
Reinforcement learning (RL) is a machine learning paradigm in which an agent learns a decision-making
policy through repeated interaction with an environment, guided by a scalar reward signal rather than
explicit supervision. me
Abstract
Reinforcement learning (RL) is a machine learning paradigm in which an agent learns a decision-making
policy through repeated interaction with an environment, guided by a scalar reward signal rather than
explicit supervision. Supervised and unsupervised learning have become indispensable tools in high-energy
physics (HEP), but they learn from fixed datasets, whereas many HEP tasks are closed-loop decisions:
the data an algorithm observes depend on the choices it makes, the feedback is sparse or delayed, and the
objective is often measured on a real machine or through a stochastic simulation rather than computed by
a differentiable model. The applications span
a gradient of maturity, from policies deployed on real accelerators and evaluated on recorded collision data
to exploratory studies in formal theory. RL is most advantageous for repeated tasks with a fast simulator,
decisions inside fast control loops, operation under faults and corrupted measurements, physics-based
rewards in place of labels, and sequential discrete constructions with a reusable policy, while Bayesian
optimization, genetic algorithms, and differentiable methods remain competitive elsewhere.Where the paper says it
Where the paper says it
Where the paper says it
This survey provides a comprehensive review of RL applications across seven major domains in HEP:
-
Accelerator Control and Beam Optimization: Covering beam steering, tuning, model-based/hybrid control, real-time feedback, and cryogenic equipment control.
-
Data Quality Monitoring (DQM): Utilizing human-in-the-loop RL and bandit algorithms to reduce expert annotation effort online and detect detector malfunctions.
-
Jet Physics: Applying RL for jet grooming, hierarchical clustering, and jet input ordering/representation.
-
Charged Particle Tracking: Formulating track finding as a graph MDP with physics-informed rewards, extended to multi-agent settings.
-
Beyond the Standard Model (BSM) and String Model Building: Exploring discrete spaces of consistent theories like flavor models and string compactifications using graph- and grammar-based RL.
-
Detector Design: Optimizing geometry and configuration using simulation-based rewards.
-
Analysis, Event Processing, and Theoretical Computation: Covering real-time trigger control, data caching for distributed analysis, systematics modeling, combinatorial event reconstruction, and formal quantum field theory (QFT) computation.
The survey discusses various RL methodologies including model-free methods (e. p. 0 This survey reviews the application of RL to such tasks across HEP: accelerator
control and feedback, data quality monitoring, jet grooming, clustering, and representation, charged
particle tracking, beyond-the-Standard-Model (BSM) and string model building, detector design, trigger
control, analysis, and formal quantum field theory computation. It provides a self-contained introduction
to the RL methods these applications rely on, covering model-free, model-based and planning, multi-agent,
offline, inverse and imitation, and bandit approaches; it reviews each domain against its traditional
methods and summarizes the available head-to-head comparisons across domains. The applications span
a gradient of maturity, from policies deployed on real accelerators and evaluated on recorded collision data
to exploratory studies in formal theory. RL is most advantageous for repeated tasks with a fast simulator,
decisions inside fast control loops, operation under faults and corrupted measurements, physics-based
rewards in place of labels, and sequential discrete constructions with a reusable policy, while Bayesian
optimization, genetic algorithms, and differentiable methods remain competitive elsewhere. 4 Policy Gradient Methods . 5 Learning in Continuous Action Spaces . This survey reviews the application of RL to such tasks across HEP: accelerator
control and feedback, data quality monitoring, jet grooming, clustering, and representation, charged
particle tracking, beyond-the-Standard-Model (BSM) and string model building, detector design, trigger
control, analysis, and formal quantum field theory computation. It provides a self-contained introduction
to the RL methods these applications rely on, covering model-free, model-based and planning, multi-agent,
offline, inverse and imitation, and bandit approaches; it reviews each domain against its traditional
methods and summarizes the available head-to-head comparisons across domains. It introduces the RL framework from
classical principles to the main deep RL algorithms, including model-based RL and planning with Monte Carlo
Tree Search. It then briefly covers the RL sub-fields most relevant to the surveyed applications: multi-agent
RL, offline RL, inverse RL and imitation learning, bandit algorithms, and, more briefly, meta, multi-task,
continual, hierarchical, Bayesian, and multi-objective RL. It then briefly covers the RL sub-fields most relevant to the surveyed applications: multi-agent
RL, offline RL, inverse RL and imitation learning, bandit algorithms, and, more briefly, meta, multi-task,
continual, hierarchical, Bayesian, and multi-objective RL. It closes with a discussion of practical challenges
that are especially relevant in the HEP setting: reward design, sim-to-real transfer, safety constraints, sample
efficiency, real-time inference and deployment, and physics-informed RL.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The paper highlights key methodological takeaways: the common first obstacle is formalizing a physics task as an MDP, requiring a shift from static learning to decision processes guided by reward functions; the necessity of embedding domain knowledge into state representation, reward design, and action space constraints; and the need for techniques like pretraining policies before fine-tuning. While RL has shown promise in matching or improving upon expert designs across many domains—such as jet grooming or track finding—it is not a universal replacement. p. 5 • Section 5: Jet physics, covering RL-based jet grooming, hierarchical clustering, and the ordering of
particle constituents for downstream classifiers. • Section 6: Charged particle tracking, where RL formulates track finding as a graph MDP with
physics-informed rewards, extended to constrained multi-agent settings. RL is most advantageous for repeated tasks with a fast simulator,
decisions inside fast control loops, operation under faults and corrupted measurements, physics-based
rewards in place of labels, and sequential discrete constructions with a reusable policy, while Bayesian
optimization, genetic algorithms, and differentiable methods remain competitive elsewhere. We conclude
with the open challenges that recur across domains, including reward design, sim-to-real transfer, safety,
the lack of shared benchmarks, and the statistical validity of policy-driven decisions.Where the paper says it
Where the paper says it
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
-
Reinforcement learning agents can be deployed in real-time feedback loops for accelerator control, achieving microsecond inference latency on dedicated hardware like FPGAs to reject disturbances and maintain stable regulation under time-varying currents.
-
Data quality monitoring systems can use bandit algorithms or human-in-the-loop RL to actively select the most informative diagnostic histograms for expert review, minimizing annotation effort compared to static threshold alarms.
-
Jet physics algorithms can utilize RL for jet grooming and clustering by learning optimal node-by-node decisions along the clustering tree based on physics rewards, improving downstream classifier performance.
-
Charged particle tracking can be formulated as a graph MDP with physics-informed rewards, enabling multi-agent RL to reconstruct tracks more effectively in complex environments like high-pileup scenarios.
-
Beyond the Standard Model theory exploration can be enhanced by using graph and grammar based RL to navigate the discrete space of symmetry groups and compactification data, leading to potentially verifiable BSM model building.
-
Detector design can be optimized using RL with simulation based rewards, allowing agents to determine the optimal geometry for calorimeters and spectrometers before physical construction.
-
Analysis systems can incorporate RL for real-time trigger control and data caching strategies, adapting thresholds during fills to manage background composition changes effectively.
Abstract
Reinforcement learning (RL) is a machine learning paradigm in which an agent learns a decision-making policy by interacting with an environment, guided by a scalar reward rather than explicit supervision. Supervised and unsupervised learning have become indispensable tools in high-energy physics (HEP), but they learn from fixed datasets, whereas many HEP tasks are closed-loop decisions: the data an algorithm observes depend on its choices, the feedback is sparse or delayed, and the objective is often measured on a real machine or in a stochastic simulation, not by a differentiable model. This survey reviews the application of RL to such tasks: accelerator control and feedback, data quality monitoring, jet grooming, clustering, and representation, charged particle tracking, beyond-the-Standard-Model (BSM) and string model building, detector design, trigger control, analysis, and formal quantum field theory computation. It provides a self-contained introduction to the underlying RL methods, from model-free and model-based to multi-agent, offline, inverse, and bandit approaches, reviews each domain against its traditional methods, and summarizes the head-to-head comparisons. The applications range from policies deployed on real accelerators and evaluated on recorded collision data to exploratory studies in formal theory. RL is most advantageous for repeated tasks with a fast simulator, decisions inside fast control loops, operation under faults and corrupted measurements, physics-based rewards in place of labels, and sequential discrete constructions with a reusable policy, while Bayesian optimization, genetic algorithms, and differentiable methods remain competitive elsewhere. We conclude with the open challenges that recur across domains, including reward design, sim-to-real transfer, safety, the lack of shared benchmarks, and the statistical validity of policy-driven decisions.
Sources
- A Living Review of Machine Learning for Particle Physics
- The Living Guide of Machine Learning for Particle Physics
- A Living Review Pipeline for AI/ML Applications in Accelerator Physics
- DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
- Deep Reinforcement Learning: An Overview
- High-Dimensional Continuous Control Using Generalized Advantage Estimation
- Proximal Policy Optimization Algorithms
- Auto-Encoding Variational Bayes
- Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems
- Safe Exploration in Continuous Action Spaces
- Trend-Based SAC Beam Control Method with Zero-Shot in Superconducting Linear Accelerator
- Optimisation of the Accelerator Control by Reinforcement Learning: A Simulation-Based Approach
- Model-free and Bayesian Ensembling Model-based Deep Reinforcement Learning for Particle Accelerator Control Demonstrated on the FERMI FEL
- Autonomous Control of a Particle Accelerator using Deep Reinforcement Learning
- Progressive Neural Networks
- Hierarchical clustering in particle physics through reinforcement learning
- Relational inductive biases, deep learning, and graph networks
- Towards Beyond Standard Model Model-Building with Reinforcement Learning on Graphs
- Lepton Flavor Portal Matter
- Autonomous Discovery of Particle Physics Theories from Experimental Data
Related papers
- DrSai: an agentic AI system toward autonomous physics analysis workflows at BESIII
- CANTON- mu Proposal: A Next-Generation Muon g-2 Measurement at Sub-0.1 ppm Precision
- Search for dark photon dark matter near 1,meV using a superconducting mixer
- ANUBIS detector and its sensitivity to neutral long-lived particles
- How to scale your HEP ML models: A recipe for robust architecture comparisons at scale
- Search for the 16 O(ppp) to 13 C pi+ pi+ e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques