From Electrons to Devices: ML-Augmented Multiscale Modeling for 2D Materials

A central challenge in computational materials science is to connect chemical-accuracy quantum mechanics with mesoscale, device-relevant simulations. This talk presents a unified, machine-learning (ML)-augmented workflow that progresses systematically from the smallest electronic scales to realistic operating conditions, while retaining rigorous physics and quantifiable uncertainty. At the quantum scale, Density-Functional Theory (DFT) simulations are benchmarked against higher-level methods to constrain ground-state energetics, magnetism/strain competition, charge-density-wave tendencies, and optical response. These results seed surrogate models that map crystal structure directly to spectra and other observables, reducing reliance on repeated many-body calculations and enabling rapid screening with calibrated error bars. At larger scales, ML interatomic potentials (ML-IAP) and data-driven kinetics capture morphology evolution and reactive environments, enabling simulations of metal nucleation and growth on 2D substrates, defect diffusion in non-stoichiometric chalcogenides, and chemo-resistive sensing under operating conditions. These mesoscale models reproduce microscopy and spectroscopy trends and are used to derive structure–property–performance relationships and operating envelopes. Transitioning to nanoscale transport, the talk examines planar MoS₂ memristors in which sulfur-vacancy density and spatial configuration govern conductance. Using DFT and Non Equilibrium Green's Functions, vacancy-dependent transmission and ON/OFF ratios are quantified. An equivariant ML-IAP, trained to DFT accuracy, drives electric-field molecular dynamics to resolve vacancy migration pathways, field-induced defect reconfiguration, and the emergence of hysteresis—linking atomistic dynamics to device-level switching and establishing design rules for stable operation. Across scales, the workflow emphasizes active-learning loops, transferability, uncertainty quantification, and modular coupling. Together these components form practical digital twins that close the gap between quantum fidelity and predictive, device-level materials simulation.