A learned force model earns trust one material at a time.
Machine learning can bring first-principles information into much larger heat-transport calculations. This perspective shows where the agreement is convincing—and why no single success validates every material or condition.
Journal of Applied Physics · 2021 · S Arabha, ZS Aghbolagh, K Ghorbani, SM Hatam-Lee, A Rajabpour
Accuracy and simulation size pull in different directions
First-principles calculations can describe atomic bonding accurately but become expensive for large or complex systems. Classical potentials scale well, yet may miss important details of a material’s energy landscape.
Where can machine-learning interatomic potentials bridge the gap between first-principles accuracy and large-scale thermal-transport simulation, and where do they still need careful validation?
Following the evidence across published studies
This perspective reviews how learned potentials are fitted to first-principles data and used in molecular dynamics and phonon-based conductivity calculations. It surveys two- and three-dimensional materials and compares published predictions with experiments, density-functional-theory-based Boltzmann transport, and classical molecular dynamics.
This is a literature perspective, not one controlled benchmark. It surveys several potential families and training strategies, then compares conductivity results obtained by molecular dynamics and Boltzmann transport. The table and figures collect values from different studies, with different cells, references, and calculation details. The useful comparison is agreement for a given system under the cited conditions, not a universal percentage advantage of machine learning.
Key findings
Training workflow
A fitted potential reduces repeated electronic-structure calculations while retaining dependence on its reference data.
2D validation
Published C₂N and silicene comparisons show useful agreement with DFT-based results, alongside differences among materials and methods.
Beyond pristine sheets
The perspective surveys bulk conductivity predictions and explains the limits imposed by training coverage and interaction complexity.
A useful bridge, with validation built in
Machine-learning potentials expand the size and time scale of atomistic thermal-transport studies, provided their training data and predictions are tested for the material and conditions of interest.
The perspective places individual successes in a broader methodological picture. It is a starting point for choosing and evaluating a learned potential for a thermal-transport problem.