Machine learning enables robust prediction of thermal boundary conductance of 2D substrate interfaces

TL;DR

This paper develops machine learning models to accurately predict the thermal boundary conductance of 2D/3D interfaces, enabling better thermal management in nanoelectronics by selecting optimal material pairings.

Contribution

It introduces ML-based predictive models for TBC of 2D/3D interfaces using first-principles data, with insights into key descriptors affecting thermal conductance.

Findings

Neural Network and Gaussian process models achieve high accuracy (RMSE < 5 MW/m²K, R² > 0.99).

Decision-tree models can predict TBC for unseen materials with RMSE < 20 MW/m²K, R² > 0.9.

Sensitivity analysis identifies key descriptors influencing TBC.

Abstract

Two-dimensional van der Waals (vdW) materials exhibit a broad palette of unique and superlative properties, including high electrical and thermal conductivities, paired with the ability to exfoliate or grow and transfer single layers onto a variety of substrates thanks to the relatively weak vdW interlayer bonding. However, the same vdW bonds also lead to relatively low thermal boundary conductance (TBC) between the 2D layer and its 3D substrate, which is the main pathway for heat removal and thermal management in devices, leading to a potential thermal bottleneck and dissipation-driven performance degradation. Here we use first-principles phonon dispersion with our 2D-3D Boltzmann phonon transport model to compute the TBC of 156 unique 2D/3D interface pairs, many of which are not available in the literature. We then employ machine learning (ML) to develop streamlined predictive models, of which Neural Network and Gaussian process display the highest predictive accuracy (RMSE $<$ 5 MWm$^{-2}$K$^{-1}$ and $R^2>$0.99) on the complete descriptor set. Then we perform sensitivity analysis to identify the most impactful descriptors, consisting of the vdW spring coupling constant, 2D thermal conductivity, ZA phonon bandwidth, the ZA phonon resonance gap, and the frequency of the first van Hove singularity or Boson peak. On that reduced set, we find that a decision-tree algorithm can make accurate predictions (RMSE $<$ 20 MWm$^{-2}$K$^{-1}$ and $R^2>$0.9) on materials it has not been trained on by performing a transferability analysis. Our model allows optimal selection of 2D-substrate pairings to maximize heat transfer and will improve thermal management in future 2D nanoelectronics.