Symmetry-Based Parameter Efficient Low-Rank Fine-Tuning for High-Throughput Calculation of Lattice Thermal Conductivity

What it claims

SBFS-FT fine-tunes MatterSim 1M (880 thousand parameters) once per crystal symmetry, using DFT data from 20 compounds. The resulting potential then computes lattice thermal conductivity (LTC) for other compounds with that symmetry, including ones it never saw. The authors are Arabov, Sozykin, Rybin and Shapeev (Skoltech, with Digital compounds LLC), and the preprint is dated October 2026. The speed-up grows with dataset size. For the 28 compounds actually tested per dataset, SBFS-FT took about 1.5 to 2 times less time than per-compound fine-tuning (SFT). The 5 to 7 times figure applies to a dataset of 100 compounds.

How it works

1. Pick 20 compounds from the same-symmetry dataset. Their average atomic masses should span roughly the dataset's whole mass range.
2. Pool about 200 to 220 DFT configurations per compound. These are displaced structures or MD frames.
3. Fine-tune MatterSim 1M for 700 epochs on the pooled set.
4. Compute LTC with Phono3py's Boltzmann transport equation in the relaxation time approximation. Cells are relaxed in ASE with FIRE and a Frechet cell filter to forces below 1e-5 eV/Å. Second- and third-order force constants use 0.03 Å displacements.

Fine-tuning is parameter-efficient. The paper compares LoRA, DoRA and LoTR. LoTR shares factor matrices across layers with a per-layer r×r core, so it scales as O(r²L + 2dr) rather than O(drL). All adapters sit only on the initial atom embedding and final output projection layers. The message-passing layers stay frozen, which the authors say prevents forgetting of pretrained physics.

The adapter method was chosen on a cheaper proxy task: phonon density of states, scored by Tanimoto similarity against MTP results on eight test compounds. The fetched text cuts off before that table, so which method won is not shown.

Evidence

The three datasets are:

• HH130: 130 half-Heusler compounds (space group 216) with MTPs.
• Spinels: 45 spinels (227) from the 4811-compound MDR database, with polyMLPs.
• 225 set: 28 compounds in space group 225, generated by the authors. They ran NVT MD with FHI-aims, then recomputed energies and forces in VASP to match MatterSim's pretraining data.

Each dataset was evaluated on 28 compounds: the 20 used for fine-tuning plus 8 held out. The metric is mean absolute relative LTC difference against the MTP or polyMLP reference, averaged over 300 to 900 K in 100 K steps. On that metric, SBFS-FT error is 2 to 3 times lower than untuned MatterSim 1M and comparable to SFT.

Limits

• The reference is LTC from MTP or polyMLP potentials, not direct DFT. Accuracy is therefore measured against another machine-learned model.
• The method still needs DFT training data for 20 compounds per symmetry.
• Each fine-tuned potential covers one space group.
• Only 8 compounds per dataset are truly held out.
• Only MatterSim 1M is tested, not the 4.5 million parameter MatterSim 5M.
• The fetched text ends partway through the results. Per-dataset error numbers, the adapter comparison, the adapter rank and any code or licence details are not shown.