Aditi Krishnapriyan

Machine learning methods for the natural sciences, generative modeling, understanding what ML models learn from scientific data, molecular and materials modeling, multi-scale dynamics, statistical mechanics, numerical methods, optimization

Our research develops machine learning methods motivated by the distinct challenges and opportunities of the natural sciences. A central theme is understanding what machine learning models learn from scientific data, rather than assuming physical structure must be hand-engineered into them, and using that understanding to develop new methods across the full arc of modeling, from architecture and training to inference-time sampling and post-training adaptation. This includes generative modeling approaches for physical systems, surrogate models for expensive simulations, methods for reliable prediction under distribution shift, and approaches that bridge the gap between simulation and experiment. Our work spans multi-scale dynamics, from quantum and atomistic systems to continuum fluids, and connects to numerical analysis, dynamical systems, statistical mechanics, quantum mechanics, and optimization. For more on our research directions and recent work, see our research page.