Accelerated and efficient modeling of low-κ organosilicate glass with the M3GNet machine learning interatomic potentials

Accurate modeling of low-κ dielectric materials is essential for advancing next-generation microelectronic devices. However, the inherent structural and chemical disorder in amorphous organosilicate glasses presents significant challenges for conventional molecular dynamics (MD) simulations, particularly when using empirical forcefields calibrated on crystalline phases. In this study, we explore the application of the Materials 3-body Graph Network (M3GNet), a machine learning interatomic potential, for high-fidelity modeling of amorphous organosilicate systems. By integrating M3GNet into extended MD workflows, we generated a chemically diverse set of low-κ organosilicate glass structures and computed their Young’s moduli. The resulting predictions of mechanical properties show excellent agreement with experimental data and outperform static deformation approaches based on density functional theory (DFT) and MD simulations using conventional force fields. These findings underscore the accuracy and transferability of M3GNet for disordered systems and demonstrate its utility in accelerating the structural and property prediction of amorphous low-κ materials. • M3GNet applied for accurate modeling of amorphous low-κ organosilicate glasses • Extended MD workflows generate chemically diverse amorphous glass structures • Young’s moduli predictions show excellent agreement with experimental data • M3GNet outperforms DFT and conventional force fields for disordered systems • Demonstrates M3GNet’s accuracy and transferability in materials property prediction