School of Physics Colloquium

Speaker: Dr. Jigyasa Nigam

Host: C. Raman

Title: Guiding insights from physics for machine-learning driven exploration of atomistic system

Abstract: Machine learning (ML)-driven computational modeling of molecules and materials has become a cornerstone of scientific inquiry, particularly in the atomic-scale search for compounds with distinct properties. Unlike many other domains, ML in this context benefits from a wealth of physical laws that govern the relationships between inputs and outputs. For example, atomic configurations, naturally represented by a set of position coordinates in 3D space, transform reliably under rotations, translations, and inversions. Most recent ML approaches which act as surrogate models of macroscopic properties, such as the potential energy surface or dipole moments, leverage domain knowledge by incorporating techniques that reflect these geometric symmetries between atomic structures and their corresponding target properties.

By embedding physical priors in intricate end-to-end architectures or designing symmetry-adapted input structural descriptors, ML has drastically accelerated the prediction, simulation, design, and characterization of diverse material systems from input geometries.

More recently, there has been a growing interest in modeling intermediate quantum mechanical (QM) components, such as electron densities, and effective single-particle Hamiltonians, which underly these structure-property relationships. With the emergence of these approaches, it has become possible to simultaneously obtain multiple output properties through established relationships or physics-based operations on these intermediate ML predictions. Given the vast design space of ML, we face a crucial question - should ML be applied directly to predict target properties, bypassing the need for QM calculations, or is its potential better realized by integrating it within a workflow that emulates QM calculations?

In this talk, I will present strategies employed by several ML frameworks to incorporate geometric symmetries and physics-based constraints. I will highlight how the integration of fundamental physical principles with data-driven methods impacts accuracy and extends the modeling capabilities to complex targets, including the self-consistent QM Hamiltonians. 

By bridging physical symmetries and the flexibility of ML, we can create models that are not only more accurate and generalizable but also transformative tools for reliable scientific exploration.

Bio: Jigyasa Nigam is a postdoc at MIT, supported by the Postdoctoral Fellowship for Excellence in Engineering. In 2024, she completed her PhD in Physics at EPFL under the mentorship of Prof. Michele Ceriotti. Her research centers on modeling molecular and material properties and the quantum mechanical workflows that underlie structure-property relationships. In particular, Jigyasa’s work emphasizes the integration of physical symmetries and constraints into machine learning models, enhancing their interpretability and reliability. Currently, she is working with Prof. Tess Smidt on investigating the robustness of equivariant models in the presence of approximate symmetries and phenomena driven by broken symmetries.