A Review of AI-Driven Approaches for Nanoscale Heat Conduction and Radiation

TL;DR

AI/ML methods are increasingly used to model nanoscale heat conduction and radiation, addressing the computational burden of DFT, MD, BTE, and RTE. The paper surveys ML prediction of phonon properties, ML interatomic potentials for MD across bulk, 2D, and interfaces, ML approaches for radiative transfer, and ML-enabled inverse design of thermal radiative devices. It identifies challenges in data availability, generalization, uncertainty quantification, and interpretability, and discusses multi-fidelity data, foundational models, and benchmark datasets as key directions. The findings show that ML surrogates can achieve large speedups (orders of magnitude) while retaining physics-informed accuracy, accelerating discovery and optimization in nanoscale thermal engineering.

Abstract

Heat conduction and radiation are two of the three fundamental modes of heat transfer, playing a critical role in a wide range of scientific and engineering applications ranging from energy systems to materials science. However, traditional physics-based simulation methods for modeling these processes often suffer from prohibitive computational costs. In recent years, the rapid advancements in Artificial Intelligence (AI) and machine learning (ML) have demonstrated remarkable potential in the modeling of nanoscale heat conduction and radiation. This review presents a comprehensive overview of recent AI-driven developments in modeling heat conduction and radiation at the nanoscale. We first discuss the ML techniques for predicting phonon properties, including phonon dispersion and scattering rates, which are foundational for determining material thermal properties. Next, we explore the role of machine-learning interatomic potentials (MLIPs) in molecular dynamics simulations and their applications to bulk materials, low-dimensional systems, and interfacial transport. We then review the ML approaches for solving radiative heat transfer problems, focusing on data-driven solutions to Maxwell's equations and the radiative transfer equation. We further discuss the ML-accelerated inverse design of radiative energy devices, including optimization-based and generative model-based methods. Finally, we discuss open challenges and future directions, including data availability, model generalization, uncertainty quantification, and interpretability. Through this survey, we aim to provide a foundational understanding of how AI techniques are reshaping thermal science and guiding future research in nanoscale heat transfer.

Figures (6)

• Figure 1: Overview of this review.
• Figure 2: ML prediction of phonon properties. (a) Virtual node GNN for predicting phonon dispersion, as adapted from Okabe et al. okabe2023virtualnodegraphneural, (b) E(3)-equivariant GNN for phonon dispersion prediction, as adapted from Fang et al. fang2024phonon, (c) ALIGNN for predicting phonon properties, as adapted from Gurunathan et al. gurunathan2023rapid, (d) Transfer learning for phonon dispersion, as adapted from Liu et al.liu2020leverage, (e) Multilayer perceptron for predicting phonon scattering rate, as adapted from Guo et al. guo2023fast, (f) random forest model for predicting phonon relaxation time, as adapted from Srivastava et al. srivastava2024accelerating, (g) Maximum likelihood estimation (MLE) method for predicting phonon relaxation time, as adapted from Guo et al. guo2023samplingacceleratedfirstprinciplespredictionphonon, (h) Combining MLE method with cutoff phonon frequency, as adapted from Zhang et al. zhang2024cryogenic.
• Figure 3: MLIP-driven predictions of thermal properties. (a) Workflow of generating a dataset using ab initio molecular dynamics simulations and training an MLIP model for property prediction, as adapted from Sours and Kulkarni Sours2023. (b) Thermal conductivity predictions for Diamond, Silicon, BAs, and InAs using MTP/ShengBTE approach, as adapted from Mortazavi et al. Mortazavi2021. (c) Illustration of training dataset for bilayer heterostructures, and the prediction accuracy of MTP for TiS2/MoS2 systems, as adapted from Nair et al. Nair2024. (d) Interfacial thermal conductance estimate from NNP-driven NEMD simulations compared to experiments and other simulation techniques, as adapted from Khot et al. Khot2025.
• Figure 4: (a) Electromagnetic spectrum from ultraviolet through microwave, highlighting the band typically considered as thermal radiation. (b) Traditional and ML/AI methods used to solve radiative heat transfer.
• Figure 5: ML prediction of radiative heat transfer. (a) Physics-informed U-net for solving Maxwell's equations, as adapted from Lim et al. Lim2022, (b) Dense NN for solving radiative transfer in participating media, as adapted from Carne et al. carne2023accelerated, (c) Dense NN for solving radiative transfer in participating media, as adapted from Stegmann et al. Stegmann2022, (d) Residual NN for solving surface-to-surface radiative transfer, as adapted from Wu et al. Wu2022, (e) CNN for dosimetry denoising, as adapted from Peng et al. peng2019cnn, (f) Tandem NN for inverse design of colored radiative cooling films, as adapted from Himes et al. Himes2022.