From QM/MM to ML/MM: A new era in multiscale modeling

Abstract

Hybrid machine-learning/molecular-mechanics (ML/MM) methods extend the classical QM/MM paradigm by replacing the quantum desription with neural network interatomic potentials trained to reproduce accurately quantum-mechanical (QM) results. By describing only the chemically active region with ML and the surrounding environment with molecular mechanics (MM), ML/MM models achieve near- QM/MM fidelity at a fraction of the computational cost, enabling routine simulation of reaction mechanisms, vibrational spectra, and binding free energies in complex biological or condensed-phase environments. The key challenge lies in coupling the ML and MM regions, task addressed through three main strategies: (1) Mechanical Embedding (ME), where ML regions interact with fixed MM charges via classical electrostatics; (2) Polarization-Corrected Mechanical Embedding (PCME), where a vacuum-trained ML potential is supplemented post hoc with electrostatic corrections, preserving transferability while approximating environment effects; and (3) Environment-Integrated Embedding (EIE), where ML potentials are trained with explicit inclusion of MM-derived fields, enhancing accuracy but requiring specialized data. Since ML/MM builds on the scaffolding of QM/MM, most proposed coupling strategies rely heavily on electrostatics, polarization, and other physicochemical concepts, the development and analysis of ML/MM schemes sits naturally at the intersection of physical chemistry and modern data science. This review surveys the conceptual foundations of ML/MM schemes, classify existing implementations, and highlights key applications and open challenges, providing a critical snapshot of the current state-of-the-art and positioning ML/MM not merely as a computational alternative, but as the natural evolution of QM/MM towards data-driven, scalable multiscale modeling.