Synthetic Training and Representation Bridging in Reconstruction Domains

TL;DR

This paper introduces a novel synthetic representation bridging the domain gap in reconstruction tasks, leveraging physical detector properties to improve interpretability and anomaly detection in high-energy physics data.

Contribution

The paper proposes a new method using synthetic data engineered to reflect detector capabilities, enhancing reconstruction accuracy and interpretability over existing latent space approaches.

Findings

Synthetic representations improve reconstruction fidelity.

Anchors neural networks to known physical methods.

Potential for better anomaly detection in physics experiments.

Abstract

Reconstructing low-dimensional truth labels from high-dimensional experimental data is a central challenge in any scenario that relies on robust mappings across this so-called domain gap, from multi-particle final states in high-energy physics to large-scale early-universe structure in cosmological surveys. We introduce a new method to bridge this domain gap with an intermediate, synthetic representation of truth that differs from methods operating purely in latent space, such as normalizing flows or invertible approaches, in that the synthetic data is specifically engineered to represent intrinsic detector hardware capabilities of the system at hand. The hypothesis is that by encoding physical properties of the detector response available only in full simulation, such synthetic representations result in a less lossy compression and recovery than a direct mapping from truth to experimental data. We demonstrate a first implementation of this concept with full simulation of a dual-readout crystal electromagnetic calorimeter for future collider detectors, in which the synthetic data is constructed to be the simulated detector hits corresponding to photon tracks of scintillation and Cerenkov photons. We refer to these signals as simulated observables as they would not be physical observables in a real detector, but are nonetheless representations of a real physical process. First results show that the synthetic representation naturally anchors the neural network architecture to a known physical method, in this case the dual-readout correction. We believe this strategy opens new avenues for machinistic interpretability and explainability of ML-based reconstruction methods. In the case of anomalous signal detection, we hypothesize that anomalous signals detected in networks trained on synthetic data rooted in a physical process are more likely to be indicative of a genuinely physical anomaly.