Digital Discovery of Large-Scale Optoelectronic Materials via MPNICE Machine-Learning Force Fields

Abstract

Accurate prediction of structure-property relationships in organic light-emitting diode (OLED) materials requires computational approaches that can efficiently capture conformational flexibility, dynamic disorder, and chemically diverse bonding environments. In this work, we demonstrate the application of machine-learning force fields (MLFFs) based on a message passing network with iterative charge equilibration (MPNICE) for predictive modeling of OLED materials. Trained on high-level electronic-structure data and incorporating iterative charge equilibration and long-range electrostatics, the MLFF framework enables rapid geometry optimization and molecular dynamics simulations with near quantum-mechanical accuracy. We show that MPNICE-optimized geometries closely reproduce density functional theory (DFT) reference structures across a diverse set of organic OLED materials, while reducing computational cost by orders of magnitude. This efficiency enables high-throughput screening workflows in which chemically enumerated molecular libraries are rapidly optimized and subsequently subjected to DFT post-processing calculations to evaluate key electronic properties.The approach is further validated through isomerization and conformational analyses, where MPNICE-based sampling accurately captures relative energetics and thermally accessible structural landscapes. Beyond static properties, we demonstrate how MPNICE-driven molecular dynamics simulations provide direct insight into the impact of molecular dynamics on optoelectronic response. For representative organometallic emitters, ensemble-averaged 1 excited-state calculations performed on snapshots extracted from MLFF trajectories yield triplet energy distributions and absorption spectra in good agreement with experimentally measured solution-phase spectra, reproducing both peak positions and spectral broadening. Collectively, these results establish MPNICE-based machine-learning force fields as a scalable and physically grounded platform for data-driven OLED materials design and accelerated optimization of next-generation optoelectronic systems.