Machine learning aided understanding and manipulating thermal transport in amorphous networks

Abstract

Thermal transport plays a pivotal role across diverse disciplines, yet the intricate relationship between amorphous network structures and thermal conductance properties remains elusive due to the absence of a reliable and comprehensive network’s dataset to be investigated. In this study, we have created a dataset comprising multiple amorphous network structures of varying sizes, generated through a combination of the node disturbance method and Delaunay triangulation, to fine-tune an initially random network toward both increased and decreased thermal conductance C. The tuning process is guided by the simulated annealing algorithm. Our findings unveil that C is inversely dependent on the normalized average shortest distance Lnorm connecting heat source nodes and sink nodes, which is determined by the network topological structure. Intuitively, the amorphous network with increased C is associated with an increased number of bonds oriented along the thermal transport direction, which shortens the heat transfer distance from the source to sink node. Conversely, thermal transport encounters impedance with an augmented number of bonds oriented perpendicular to the thermal transport direction, which is demonstrated by the increased Lnorm. This relationship can be described by a power law C=Lnormα, applicable to the diverse-sized amorphous networks we have investigated.