Rational Design of Light-Driven Molecular Motors Enabled by a Data-Efficient Machine Learning Framework for Nonadiabatic Dynamics

Abstract

Light-driven molecular rotary motors convert photon energy into mechanical motion, but designing systems that combine high rotational speed with robust unidirectionality remains a fundamental challenge. Conventional four-stroke motors rely on thermally activated steps that often slow complete rotary cycles to microsecond-to-second or even longer time scales, whereas emerging bioinspired designs may achieve much faster photoinduced rotation. However, the computational study of both classes is limited by the prohibitive cost of quantum-chemical nonadiabatic dynamics over the extended time scales required to resolve productive rotary motion. Here, we introduce N2ZND, a Neural Network Zhu–Nakamura Dynamics framework that combines hierarchical sampling with E(3)-equivariant deep learning to enable data-efficient simulations of excited-state molecular motor dynamics. Trained on sparse, chemically targeted high-level ab initio data, N2ZND achieves chemical accuracy and enables extensive nonadiabatic simulations of light-driven motors across relevant time scales. Application of N2ZND to a complex bioinspired motor reveals a two-stroke operating mechanism that bypasses the thermally activated barriers characteristic of conventional four-stroke motors, enabling complete rotary motion on a picosecond time scale. We further identify an intrinsic speed-directionality trade-off: removal of thermal gating accelerates rotation but can reduce directional fidelity through excited-state wavepacket bifurcation. Guided by this mechanistic insight, we rationally designed a structurally modified motor in which steric tuning reshapes the excited-state topology, restoring robust room-temperature unidirectionality while preserving picosecond rotary operation. These results establish N2ZND as a scalable platform for mechanistic discovery and rational design of light-driven molecular machines, providing a foundation for future high-throughput screening across traditional and bioinspired motor architectures.