Machine Learning-driven Discovery of New Thermal Transport Mechanisms in Porous Materials

Han Wei¹, Hua Bao^1*, Xiulin Ruan^2*

¹ University of Michigan-Shanghai Jiao Tong University Joint Institute, Shanghai Jiao Tong University, Shanghai 200240, China；

² School of Mechanical Engineering and Birck Nanotechnology Center, Purdue University, West Lafayette, Indiana 47907, USA

ABSTRACT: In material science, machine-learning algorithms have been successfully applied to predict material properties, but they are rarely used to uncover new physical mechanisms. In this work, we aim to use machine learning to facilitate discoveries of exceptions beyond accepted knowledge and demonstrate the new research paradigms that promote the development of theoretical studies. Manipulating the thermal transport in nanostructured materials is of great significance for promoting heat dissipation in electronic chips and improving the energy conversion efficiency of thermoelectric materials. Previous studies have found that in nanoporous materials, random pores can lead to reduced thermal conductivity than the periodic pore arrangement. This is because the random arrangement induces the localization of heat carriers, phonons, thereby hindering the transfer of heat. In this work, we investigated the thermal transport in nanoporous graphene with disordered pore arrangements. First, we attempted a manual search for random structures with possibly higher thermal conductivities than the periodic structure. Molecular dynamics simulations were performed to calculate the thermal conductivities of 21 random porous graphene structures, which are all lower than the periodically arranged one. To search more efficiently, we proposed a "two-step" search method based on genetic algorithm. During genetic algorithm optimization, thermal conductivities of candidates were computed by solving the Boltzmann transport equation, which is a less computationally expensive model. BTE simulation is carried out for pre-screening and then NEMD simulation is performed to validate the optimal configuration found by genetic algorithm. The total computational cost needs to be highlighted. In the unsuccessful manual search, about 67200 core hours were spent, while no disorder-enhanced case is found. For the successful genetic algorithm enabled search, with only 6604 core hours, two disorder-enhanced cases were found. Therefore, this genetic algorithm enabled two-step search strategy largely reduces the computational cost while increases the probability of success.

Through an in-depth analysis of the characteristics of the porous structure with high thermal conductivity, it is found that the enhancement of thermal conductivity of disordered configurations was attributed to the particular pore arrangements. Inspired by observation, we proposed two structural parameters—shape factor and channel factor, to describe the characteristics of pore arrangement. Further study shows that the shape factor and the channel factor play an important role in the heat conduction of porous graphene. Their effects far exceed the phonon localization effect, which leads to the enhancement of thermal conductivity. Further regression analysis shows that the shape factor and channel factor have strong positive correlations with thermal conductivity. These two new physical parameters can be used to establish a prediction model for fast predicting the thermal conductivity of nanoporous structures.

Besides nanoporous materials, the microscopic morphology of the pores will also significantly affect the heat transport of the porous media in the macro scale. However, existing effective medium approaches usually miss morphology effects. Although numerical simulation can capture the structural information, they are expensive and not physically intuitive. Machine learning methods have recently been successful in predicting effective thermal conductivity, but the lack of descriptors limits physical insights. In order to accurately predict the thermal conductivity of porous media and at the same time guide the understanding of heat transfer mechanism, developing physics-based descriptors to quantify the complex morphologies in porous media is highly desirable. Therefore, besides shape factor and channel factor, we proposed other three structural descriptors to characterize the structural characteristics of porous media: bottleneck, perpendicular nonuniformity, and dominant paths. The anisotropy of the shape and distribution of pores can significantly affect the effective thermal conductivity of porous media, which can be effectively quantified by these descriptors. They have strong correlations with thermal conductivities and the machine learning model based on these descriptors can accurately predict the thermal conductivity of porous media with complex morphology.

Keywords: machine learning; genetic algorithm; optimization; porous media; effective thermal conductivity; structural descriptor