AI-Accelerated Atom-to-Device Thermal Modeling of Nanoscale Semiconductor Devices

The Apple A18 chip in the iPhone 16e integrates ~15 billion transistors fabricated using a 3 nanometer process, where each nanometer-scale transistor comprises heterogeneous semiconductor, dielectric, and metallic layers confined within extremely small dimensions. Such confined geometries lead to localized heating that accelerates defect generation and device degradation. At these scales, heat conduction deviates fundamentally from bulk behavior due to phonon confinement and boundary scattering, rendering conventional Fourier-based models inadequate. While atomistic modelling can accurately predict thermal properties of nanoscale systems, the computational expense of first-principles approaches prevents their direct application to realistic device geometries. In this talk, I will present a framework that integrates atomistic modelling with machine-learning (ML) methods to overcome these challenges and predict the thermal properties of sub-10 nm field-effect transistors. We construct atom-by-atom digital twins of multi-interface transistor heterostructures and compute their thermal behavior using Graphics Processing Units Molecular Dynamics (GPUMD), a fully GPU-accelerated simulation package. The GPUMD framework is used to train ML interatomic potentials, neuroevolution potentials (NEPs), which, combined with MD simulations, enable efficient calculation of in-plane and cross-plane thermal conductivities, interfacial thermal conductance, and layer-specific thermal resistances under both steady-state and transient heating. The resulting atom-to-device thermal model provides physics-based insights to guide the design of nanoscale transistors with targeted thermal properties.