Multiscale multiphysics simulation and machine learning for energy transport

<p dir="ltr">Efficient energy utilization is essential for sustainable development and combating global energy crisis, yet a substantial portion of energy is lost as waste heat due to limitations in thermal management. To address this challenge, this dissertation investigates energy transport across multiple scales, with a focus on lattice thermal transport and radiative cooling. By combining multiphysics modeling, machine learning (ML), and high-performance computing (HPC), this dissertation aims to advance our understanding and predictive capabilities for thermal transport phenomena in complex materials and devices. </p><p dir="ltr">At the atomic scale, phonon-phonon scattering dominates thermal transport in semiconductors and insulators. While first-principles methods based on density functional theory and the Boltzmann transport equation offer accurate predictions, their computational cost becomes prohibitive, especially when including four-phonon scattering. This dissertation focuses on accelerating the phonon scattering simulations. We first introduce a neural network model to directly predict all phonon scattering rates, which achieves up to two orders of magnitude speedup with near-first-principles accuracy. Transfer learning methods are further developed to improve the model’s performance. Inspired by this work, we then developed a sampling and maximum likelihood approximation framework to reduce the computational cost. By only analytically calculating a small portion of scattering processes, the simulation is accelerated even more, by three to four orders of magnitude. Additionally, a GPU-accelerated computing framework is implemented to support large-scale simulations without sacrificing precision. The method leverages the strong parallel computing ability of GPU to accelerate the simulation, offering over ten times speedup. These methods reduce the computational overhead of studying complex material or doing high-throughput screening of thermal properties. </p><p dir="ltr">Next, we apply these advances to study anisotropic thermal transport in hexagonal boron nitride (h-BN), a 2D material with high thermal conductivity and unique dielectric properties. Leveraging our sampling-based acceleration method, we employ a dense q-mesh to achieve fully converged simulations. By incorporating three-phonon, four-phonon scattering, and phonon renormalization, our simulations reveal the competing effects that govern h-BN’s anisotropic thermal conductivity and zone-center phonon linewidths, achieving excellent agreement with experimental results. </p><p dir="ltr">Building on our understanding of thermal transport at the atomic scale, this dissertation extends to the nanoscale by investigating the radiative cooling applications of h-BN-based coatings. A multiscale multiphysics framework is constructed by coupling first-principles optical calculations, finite element analysis, and Monte Carlo photon transport, revealing the atomic and structural origins of high solar reflectance and reduced mid-IR emissivity. While the ultra-white radiative cooling materials can achieve a high solar reflectance, color paints are more desirable for everyday applications due to aesthetic considerations.</p><p dir="ltr">In Chapter 7 , we focus on the practical challenges in the design of colored radiative cooling paint. Conventional, intuition-driven methods rely on trial and error and often fail to achieve optimal thermal performance. By integrating optimization techniques, photon Monte-Carlo simulations, and ML models, we achieve an inverse design of colored radiative cooling paint that tailors coatings to a target color while maximizing cooling performance. Our framework enables the practical inverse design of high-performance colored radiative cooling paints. </p><p dir="ltr">Together, these contributions provide a unified computational strategy for understanding, predicting, and designing energy transport materials. The integration of ML, HPC, and multiscale physics in this dissertation offers new pathways for developing next-generation thermal management and radiative cooling technologies.</p>