End-to-end Material Thermal Conductivity Prediction through Machine Learning

For the particular case of thermal transport, while these approaches are gaining popularity, they are still limited. Thermal conductivity (κ) is an important material For instance, Pal et al. [25] employed a scale-invariant property critical in determining the performance and efficiency ML model to accelerate the search of quaternary chalcogenides of devices in various technological applications with low κ, Hu et al. [26] employed ML to minimize such as thermoelectric energy generation, thermal insulation, coherent heat conduction across aperiodic superlattices, and memory storage [1-4]. For many of these applications, Rodiguez et al. [27] trained neural network based low thermal conductivity semiconducting solids interatomic forcefield to do bottom-up prediction of κ are desired, while for others (such as heat dissipation based on intermediate phonon properties such as mean and microprocessors), materials with high κ are desired square displacements and bonding/anti-bonding characters, [2, 5, 6]. For materials used in most of these applications, and Visaria and Jain [28] employed neural network the thermal transport is dominated by atomic vibrations, based auto-encoders to do space transformation to search i.e., phonons, with room temperature κ in the range of for material configurations with low-and high-κ from the 0.1-3000 W/m-K [7]. The traditional search for novel low exponentially-large search space of considered superlattices.