E.M. Collins, K.B. Moore, H. Abroshan, D. Giesen, M.D. Halls, A. Chandrasekaran
Schrödinger, Inc.,
United States
Keywords: machine learning, quantitative structure-property relationships, formulations, optoelectronics, molecular quantum mechanical calculations
Summary:
The simulation of material properties using physics-based approaches, such as density functional theory (DFT) and time-dependent DFT (TD-DFT), has proven invaluable in understanding structure-property relationships and guiding materials design. While these methods offer powerful insights, they face inherent limitations in computational scaling and cost, particularly for large-scale materials screening. Machine learning (ML) has emerged as a promising complement to traditional physics-based modeling, offering the potential to dramatically accelerate materials innovation while maintaining physical accuracy. In this talk, we first demonstrate how combining ML with physics-based approaches can overcome these challenges in designing functional materials, such as battery electrolytes, organic light-emitting diodes (OLEDs), and fluorescent dyes. By incorporating physical insights into our ML frameworks, we show how these hybrid approaches can maintain accuracy even in data-limited regimes while significantly improving computational efficiency. We then explore the extension of these methods to more complex systems, particularly formulations or mixtures of multiple materials, where emergent properties arise from subtle intermolecular interactions dependent on both structure and composition. Through the evaluation of various molecular representations and ML architectures, we demonstrate strategies for optimizing both predictive power and computational throughput. Finally, we showcase how these developed frameworks can be applied to accelerate the discovery and design of novel materials with targeted properties. This work highlights the potential of combining physics-based modeling with machine learning to advance materials innovation across multiple domains.