Differentiable Hybrid Machine Learning Framework for Nanoscale Thermal Transport

The semiconductor industry has experienced significant progress in recent decades, driving the demand for faster and more energy-efficient electronic devices. This development necessitates the miniaturization of transistors, leading to elevated hot spot temperatures, which can adversely affect device longevity and reliability. Accurately predicting these temperatures is essential for transistor design optimization. At the nanoscale, traditional heat transfer models like Fourier's law are inadequate, and the Phonon Boltzmann Transport Equation (BTE) becomes essential for precise temperature profiling. However, BTE's nonlinear and high-dimensional nature poses substantial challenges for traditional numerical solvers, especially when dealing with complex heat transfer in multi-material systems involving different geometries and interfaces. To address these challenges, this talk introduces a novel approach utilizing a physics-integrated differentiable programming framework. This model is designed to preserve the mathematical structures inherent in the physics of BTE and the associated boundary conditions at interfaces. The framework effectively integrates machine learning with physical laws, creating a hybrid model that not only navigates the complexities of BTE but also allows for solving inverse problems. This approach represents a significant advancement in the field of nanoscale heat transfer, demonstrating the potential of machine learning to solve complex, high-dimensional thermal transport problems.