Machine Learning-Based Classifiation of Nuclear Equations of State Using Sup ernova Gravitational Wave Data

Abstract

This thesis is fo cused on the explainability of machine learning mo dels in the context of nuclear equation of state classifiation. Can a CNN identify the nuclear e quati on of state (EOS) from a core-collapse sup ernova gravitational wave signal? And are the features it uses actually physics? Two pip elines were trained on 664 simulated GW signals (SFHo, LS220, DD2, GShenFSU2.1): a 1D CNN on raw − 2 to 6 ms waveforms, reaching 97.74% accuracy, and a WaveletCNN on 128 × 328 Morlet scalograms at 86.47%. The accuracy is a prerequisite. The main result is about what drove those decisions. Six XAI methods (gradient saliency maps, temporal masking, Grad-CAM++, Guided Grad-CAM, occlusion sensitivity, and DeconvNet) all identify the 1.5–2.5 ms post-bounce ring-down in the 600–1800 Hz band as the discriminative signal. Masking that window drops accuracy by 62 points. The pre-bounce and late-tail (>4 ms) regions contribute nothing. Aggregated Grad-CAM++ and occlusion maps confim the result across the full validation population, not just individual examples. Agreement across methods with diffrent computational mechanisms rules out a shared simulation artefact; what they agree on is the protoneutron star oscillation. Both classifirs are validated against known PNS 𝑓mode predictions. The framework is applicable to future Galactic supernova detections by Advanced LIGO, Virgo, and KAGRA