Optimization of Thermal Conductivity at Interfaces Using Learning Algorithms

In material science, we are frequently interested in understanding the properties and design implications of material at interfaces. These interfaces can be manipulated to improve the desired characteristics of the bulk material. Optimization of thermal transport across the interface of two different materials is critical to micro/nanoscale electronic, photonic, and phononic devices. However, interfacial thermal conductivity optimization is a high-dimensional problem due to the number of factors affecting it. Non-equilibrium molecular dynamics (NEMD) has been used to investigate the thermal transport across a solid-solid interface between different materials. Despite NEMD being an effective approach to this problem, scanning the large parameter space is unrealistic due to high computational cost. Alternatively, machine learning is a cost-effective and time-efficient method that can investigate this interfacial thermal transport phenomenon.

In this talk, we describe our efforts to understand and optimize the impact of interfacial atomic defects and intermixing on the thermal transport across an Al/Si junction. We developed a learning-based framework to optimize over a parameter space that would be prohibitively large in the brute force approach. Using this technique allows us to accumulate knowledge of the system of a given type of atoms and store this information into a neural network. A scalable framework was built to allow utilization of the available high-performance computing resources, thereby accelerating the learning by directly interfacing with the LAMMPS molecular dynamics simulation package. We found that mixing 3-4 monolayer of atoms near the interface can significantly increase the thermal conductance. However, when inter atomic mixing layers get thicker, it induces a negative effect on the thermal conductance due to phonon scattering. The model is able to find the optimal mixing length and mixing fraction relatively fast. This work can be extended to more systems by inclusion of descriptors and a high-fidelity surrogate model.