Enhancing High-Dimensional Neural Network Potentials Accuracy in OLED Systems via Element-Relabeling

Abstract

Accurate atomistic simulations are essential for understanding organic light-emitting diode (OLED) materials in complex condensed-phase environments. However, ab initio methods are computationally demanding, whereas classical force fields often lack sufficient accuracy to represent complex molecular behavior. In this work, we present a high-dimensional neural network potential (HDNNP) framework specifically optimized for complex OLED systems. Detailed atomic-level analysis reveals that local environments, measured by intramolecular neighbor density, serve as a key factor of force prediction heterogeneity. To reduce force prediction errors arising from intrinsic molecular complexity, we introduce a conceptually simple yet effective strategy: relabeling atoms exhibiting large force errors as distinct pseudo-elements. By providing more environment-specific local descriptors, this targeted refinement improves force and total-energy prediction accuracy. Such accuracy gains lead to higher fidelity in the reproduction of key simulation observables. To demonstrate the transferability of the element-relabeling strategy, we applied it to chemically diverse benchmark systems, specifically aspirin and doublewalled carbon nanotube. Our approach substantially reduces the accuracy gap with state-of-the-art graph-based models while preserving the computational efficiency of conventional HDNNPs, enabling large scale simulations, exemplified by accurate glass transition temperature prediction.