SAMRAT
CHAKRABORTY

I research interpretable machine learning and ML force fields (MLFFs) to probe collective physical phenomena that black-box models obscure and classical methods fail to quantify. Under Dr. Avisek Das at IACS, I developed a 3-tier interpretability framework (Descriptor Informativeness Index [DII], combinatorial ablation, and Shapley attribution) across system-tailored physical coordinate sets (ranging from 20 to 60+ descriptors depending on mixture complexity) to uncover the collective variables governing non-covalent bond lifetimes.

To make rigorous physical chemistry accessible without GPU cloud lock-in, I engineered a high-performance streaming trajectory engine with capped neighbor-lists that processes 200 ns runs of 100 Å bulk water boxes in under 10 minutes on <1.2 GB RAM (an 800× speedup over default MDAnalysis), while enabling surrogate models of ~5,000 water molecules to train on CPU in under 17 minutes. Across domains, my research is unified by a single conviction: models are mathematical probes to isolate physical causality, not black-box predictors. Whether formulating calibration-gated active learning metrics (ROE) for self-driving laboratories or architecting autonomous closed-loop telemetry systems (SATISH), I build systems where every prediction is physically grounded, auditable, and interpretable.