RESuM: Rare Event Surrogate Model for Physics Detector Design

Ann-Kathrin Schuetz, A. W. P. Poon, A. LiarXiv 2024 · DOI · arXiv

Abstract

The experimental discovery of neutrinoless double-beta decay (NLDBD) would answer one of the most important questions in physics: Why is there more matter than antimatter in our universe? To maximize the chances of detection, NLDBD experiments must optimize their detector designs to minimize the probability of background events contaminating the detector. Given that this probability is inherently low, design optimization either requires extremely costly simulations to generate sufficient background counts or contending with significant variance. In this work, we formalize this dilemma as a Rare Event Design (RED) problem: identifying optimal design parameters when the design metric to be minimized is inherently small. We then designed the Rare Event Surrogate Model (RESuM) for physics detector design optimization under RED conditions. RESuM uses a pretrained Conditional Neural Process (CNP) model to incorporate additional prior knowledges into a Multi-Fidelity Gaussian Process model. We applied RESuM to optimize neutron moderator designs for the LEGEND NLDBD experiment, identifying an optimal design that reduces neutron background by ($66.5\pm3.5$)% while using only 3.3% of the computational resources compared to traditional methods. Given the prevalence of RED problems in other fields of physical sciences, the RESuM algorithm has broad potential for simulation-intensive applications.

Introduction: The Rare Event Design Dilemma

In the search for fundamental physics phenomena like neutrinoless double-beta decay (NLDBD), detectors must achieve unprecedented levels of "quietness." A single background event, such as a stray neutron, occurring once every few years can compromise an entire experiment. Designing these detectors requires optimizing complex geometries to minimize these rare background events.

Traditionally, scientists rely on Monte Carlo simulations to evaluate designs. However, they face the Rare Event Design (RED) dilemma: because background events are extremely rare, obtaining statistically significant counts requires millions of expensive simulations. Alternatively, researchers must work with "noisy" data—mostly zeros and ones—that standard optimization algorithms struggle to process. This makes finding optimal detector geometry a slow and computationally prohibitive task.

Methodology: The RESuM Framework

The authors propose RESuM (Rare Event Surrogate Model), a three-layered approach designed to overcome the RED problem by integrating event-level information into the optimization process.

The RESuM framework architecture combining CNP and MFGP.
  1. Conditional Neural Process (CNP): The core innovation of RESuM. Instead of aggregating simulation results into a final count, the CNP analyzes event-level data. It learns to assign a continuous "score" ($\beta$) to every simulated particle, transforming binary outcomes ("hit" or "miss") into a smooth probability landscape. This significantly reduces the variance associated with discrete Poisson counts.
  2. Multi-Fidelity Gaussian Process (MFGP): The framework utilizes two simulation tiers. Low-fidelity (LF) simulations use simplified physics and are over 1,000 times faster, enabling broad exploration. High-fidelity (HF) simulations provide accuracy but are computationally expensive. The MFGP learns the correlation between these fidelities, using cheap data to guide the placement of expensive simulations.
  3. Active Learning: RESuM employs an acquisition function to intelligently select the next design parameters for simulation, prioritizing areas of high uncertainty to maximize information gain.

Experimental Setup and Datasets

The study focuses on the LEGEND experiment, a leading search for NLDBD.

  • Training Data: 304 low-fidelity simulations (0.15 CPU hours each) and 10 high-fidelity simulations (170 CPU hours each).
  • Validation Data: 100 independent high-fidelity simulations were used to verify model predictions.
  • Physics Engine: Data was generated using the GEANT-4 toolkit, the industry standard for simulating particle interactions with matter.

Key Insights: Smoothing the Discrete

The primary breakthrough is using the CNP to bypass the inherent noise of rare events. By predicting a continuous probability score for every particle, the model creates a differentiable surface suitable for Gaussian Process modeling.

Comparison of raw simulation counts vs. CNP-smoothed scores across design parameters.

To handle the massive class imbalance—where only 1 in 50,000 particles might be a "signal" event—the authors utilized Mixup, a data augmentation technique that creates synthetic training examples by blending signal and background events. This allows the CNP to act as a mechanism for injecting event-level prior knowledge into the trial-level surrogate model, drastically improving data efficiency.

Key Results

  • Computational Efficiency: RESuM identified optimal designs using only 1,746.5 CPU hours. In contrast, traditional grid-search methods would have required approximately 52,700 CPU hours to reach the same resolution—a 96.7% reduction in cost.
  • Optimized Design: The model discovered a neutron moderator configuration that reduces background by (66.5 ± 3.5)% compared to the baseline design.
  • Reliable Uncertainty Quantification: Validation tests confirmed that the model's uncertainty bands accurately reflect reality, with 95% of ground-truth data points falling within the 2$\sigma$ confidence interval.

Conclusion and Future Directions

While the results are promising, the authors note limitations, such as the small validation sample size (100) necessitated by the high cost of HF simulations. Future work could involve more sophisticated active learning strategies to locate global optima even more rapidly.

Beyond particle physics, RESuM has potential applications in any field where simulation-intensive optimization meets rare-event metrics, such as modeling Binary Black Hole mergers in astronomy or exploring molecular structures via Density Functional Theory in material science.

Summary updated

Abstract

The experimental discovery of neutrinoless double-beta decay (NLDBD) would answer one of the most important questions in physics: Why is there more matter than antimatter in our universe? To maximize the chances of detection, NLDBD experiments must optimize their detector designs to minimize the probability of background events contaminating the detector. Given that this probability is inherently low, design optimization either requires extremely costly simulations to generate sufficient background counts or contending with significant variance. In this work, we formalize this dilemma as a Rare Event Design (RED) problem: identifying optimal design parameters when the design metric to be minimized is inherently small. We then designed the Rare Event Surrogate Model (RESuM) for physics detector design optimization under RED conditions. RESuM uses a pretrained Conditional Neural Process (CNP) model to incorporate additional prior knowledges into a Multi-Fidelity Gaussian Process model. We applied RESuM to optimize neutron moderator designs for the LEGEND NLDBD experiment, identifying an optimal design that reduces neutron background by ($66.5\pm3.5$)% while using only 3.3% of the computational resources compared to traditional methods. Given the prevalence of RED problems in other fields of physical sciences, the RESuM algorithm has broad potential for simulation-intensive applications.
BibTeX
@misc{schuetz2024resum,
  author = {Ann-Kathrin Schuetz and A. W. P. Poon and A. Li},
  title = {{RESuM: Rare Event Surrogate Model for Physics Detector Design}},
  year = {2024},
  eprint = {2410.03873},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.03873},
}
Figure