Symmetry-Based Parameter Efficient Low-Rank Fine-Tuning for High-Throughput Calculation of Lattice Thermal Conductivity
Abstract
Accurate and efficient prediction of lattice thermal conductivity (LTC) is essential for developing materials for thermoelectric, microelectronic, and photovoltaic applications. Here, we present a methodology for high-throughput calculation of LTC across large datasets of compounds sharing a common crystal symmetry. Our approach combines fine-tuning of the MatterSim 1M universal machine-learned interatomic potential with LTC calculations based on the solution of the Boltzmann transport equation. We introduce a symmetry-based few-shot fine-tuning strategy (SBFS-FT) based on low-rank adaptation. In this approach, MatterSim 1M is fine-tuned on a small, carefully selected set of compounds that share a common crystal symmetry. The resulting potential is then applicable for accurate LTC calculations for compounds with this crystal symmetry, even if they were not included in the fine-tuning set. We tested our methodology on datasets of half-Heusler compounds (space group 216), materials with space group 225, and spinels (space group 227). We demonstrated that LTC calculations with MatterSim 1M fine-tuned using SBFS-FT exhibited a mean absolute relative difference from the reference results (obtained using moment tensor potentials and polyMLPs) 2-3 times smaller than that achieved with the non-fine-tuned MatterSim 1M potential. Moreover, the accuracy of LTC calculations with MatterSim 1M after SBFS-FT was comparable to that achieved using the MatterSim 1M potentials individually fine-tuned for each compound. Importantly, SBFS-FT enabled generation of the potential suitable for accurate LTC calculations across large compound datasets substantially faster than fine-tuning of MatterSim 1M for each compound individually, reducing the computational time by approximately a factor of 5-7 for a dataset of 100 compounds.
Related
- High-throughput study of electrical conductivity in ordered metals
- Efficient calculation of inductive coupling for arrays of wire ring resonators
- Calculation of structural symmetry and applicability of small AkBlCm nanoclusters (A=Rb, Cs; B=Pb, Sn, Bi; C=Cl, Br, I) as building blocks for synthesis of larger-size nanoclusters
- Accurate and efficient calculation of atomic forces in solids with non-self-consistent hybrid functionals
- Efficient Calculation of Equilibrium Correlation Functions
BibTeX
@misc{arabov2026symmetrybased,
title = {{Symmetry-Based Parameter Efficient Low-Rank Fine-Tuning for High-Throughput Calculation of Lattice Thermal Conductivity}},
author = {Rustam Arabov and Konstantin Sozykin and Alexander Shapeev and Nikita Rybin},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.29634},
eprint = {2610.29634},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.29634},
}
RIS
TY - UNPB
TI - Symmetry-Based Parameter Efficient Low-Rank Fine-Tuning for High-Throughput Calculation of Lattice Thermal Conductivity
AU - Rustam Arabov
AU - Konstantin Sozykin
AU - Alexander Shapeev
AU - Nikita Rybin
PY - 2026
DA - 2026/10/27
PB - arXiv
DO - 10.48550/arxiv.2610.29634
UR - https://arxiv.org/abs/2610.29634
AB - Accurate and efficient prediction of lattice thermal conductivity (LTC) is essential for developing materials for thermoelectric, microelectronic, and photovoltaic applications. Here, we present a methodology for high-throughput calculation of LTC across large datasets of compounds sharing a common crystal symmetry. Our approach combines fine-tuning of the MatterSim 1M universal machine-learned interatomic potential with LTC calculations based on the solution of the Boltzmann transport equation. We introduce a symmetry-based few-shot fine-tuning strategy (SBFS-FT) based on low-rank adaptation. In this approach, MatterSim 1M is fine-tuned on a small, carefully selected set of compounds that share a common crystal symmetry. The resulting potential is then applicable for accurate LTC calculations for compounds with this crystal symmetry, even if they were not included in the fine-tuning set. We tested our methodology on datasets of half-Heusler compounds (space group 216), materials with space group 225, and spinels (space group 227). We demonstrated that LTC calculations with MatterSim 1M fine-tuned using SBFS-FT exhibited a mean absolute relative difference from the reference results (obtained using moment tensor potentials and polyMLPs) 2-3 times smaller than that achieved with the non-fine-tuned MatterSim 1M potential. Moreover, the accuracy of LTC calculations with MatterSim 1M after SBFS-FT was comparable to that achieved using the MatterSim 1M potentials individually fine-tuned for each compound. Importantly, SBFS-FT enabled generation of the potential suitable for accurate LTC calculations across large compound datasets substantially faster than fine-tuning of MatterSim 1M for each compound individually, reducing the computational time by approximately a factor of 5-7 for a dataset of 100 compounds.
ER -