SINAPSE introduces lightweight explainable neutron-gamma discrimination

According to the arXiv submission 2605.13627 (submitted 13 May 2026), Thomas Carreau and 13 coauthors introduce SINAPSE, described as a lightweight deep learning framework for accurate and explainable neutron-gamma discrimination in organic scintillators. The paper reports a dual-branch architecture that combines a 1-dimensional convolutional autoencoder for waveform denoising with a downstream classifier. The authors report applying random augmentations to high-SNR waveforms to emulate low-charge, low-SNR conditions, and claim SINAPSE delivers superior denoising vs conventional digital signal processing techniques and yields well-calibrated output probabilities (per the manuscript). The paper also reports using SHAP (SHapley Additive exPlanations) to attribute model decisions to pulse-shape features consistent with established pulse-shape discrimination (PSD) principles.

Researchers working on detector-level signal classification commonly use denoising autoencoders and joint denoise-classify architectures to improve signal fidelity before label-driven inference. Industry-pattern observations: applying random augmentation on high-quality waveforms to simulate low-charge regimes is a pragmatic approach to extend supervised models into data-poor SNR regimes without relying solely on unreliable PSD labels. Using SHAP for feature attribution follows a growing practice of combining model-level performance metrics with post hoc explainability to validate that learned features map to known physics.

For experimental instrumentation and applied ML teams, the paper demonstrates an integration of denoising, calibration of output probabilities, and explainability in a single lightweight pipeline. Industry-pattern observations: similar pipelines can reduce dependence on handcrafted graphical cuts in low-charge regimes, but they require careful benchmarking against conventional DSP across SNR ranges and attention to calibration under domain shift.

Observers should look for an open-source release of the code and trained weights, quantitative benchmark tables comparing SINAPSE to standard PSD metrics across SNR bins, latency and memory footprints for on-detector or FPGA/edge deployment, and validations on independent experimental datasets. The arXiv submission provides the design and initial claims; replication and public benchmarks will determine practical adoption.