Explainable AI for Binary Black Hole Light Curve Classification via Feature-Weighted Embeddings

Classifying binary black-hole (BBH) systems from noisy, irregular light curves poses a significant challenge in astrophysics. This work introduces an explainable AI pipeline designed to overcome this roadblock by comparing accretion rate time series from a target source with a catalog of simulations. Our methodological approach involves three key stages: (i) automated feature extraction of accretion-rate time series using pycatch22; (ii) supervised feature-weight learning via multi-output Random Forests to generate interpretable importance scores; and (iii) nearest-neighbor retrieval in a weighted-embedding space for inferring the mass ratio (q) and eccentricity (e) of the BBHs. This pipeline effectively leverages the scalability of modern machine learning techniques while retaining crucial physical interpretability. We present preliminary results using 220 segments of simulated time series, that show that our method achieves over 90% accuracy in recovering the correct (q, e) combination, even when including observational sampling and noise.