Light-driven molecular nanomotors (MNMs) have attracted growing interest in the biomedical field due to their promising applications in targeted drug delivery, membrane permeation. The operational efficiency of these systems is governed by their ability to undergo controlled unidirectional rotation, which is fundamentally determined by their excited-state photophysics. Accurate photophysical characterization typically requires high-level quantum chemical calculations, including potential energy surfaces, nuclear forces, and non-adiabatic coupling vectors. However, the high computational cost of such methods severely limits their applicability to large systems and long-time nonadiabatic molecular dynamics simulations. In this work, we explore SPaiNN, a machine-learning framework based on equivariant message-passing neural networks, as an efficient surrogate model for excited-state properties. Trained on quantum chemical reference data, SPaiNN predicts energies, forces, and non-adiabatic couplings for multiple electronic states, enabling computationally efficient nonadiabatic molecular dynamics simulations of molecular nanomotors.
Pragati Rohali