Systematic softening in universal machine learning interatomic potentials - Fine-tuning

Aug 3, 2026

Upload by Liam

Supplementary material for addressing systematic softening in universal machine learning interatomic potentials through fine-tuning. The content provides benchmark and diagnostic sections including surface energy evaluation with explicit Miller indices and MPIDs, followed by defect benchmarks, phonon vibrational frequencies, ion migration barrier analysis, and discussion of softening scales. It also compares fine-tuning against training from scratch and reports additional experiments such as linear corrected CHGNet for surface calculation and the effect of model size.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages27

Reading Time1 h 8 min

AccessFree to download

Surface Benchmark

Defect Benchmark

Phonon Vibrational Frequencies

Ion Migration Barriers

Softening scales

Fine-tuning vs Training from-scratch

Linear corrected CHGNet for surface calculation

Effect of model size

References

Supplementary Information for overcoming systematic softening in universal machine learning interatomic potentials by fine-tuning CONTENTS I. Surface Benchmark 2 II. Defect Benchmark 8 III. Phonon Vibrational Frequencies 13 IV. Ion Migration Barriers 21 V. Softening scales 21 VI. Fine-tuning vs Training from-scratch 21 VII. Linear corrected CHGNet for surface calculation 24 VIII. Effect of model size 25 References 26 I. SURFACE BENCHMARK The table below presents all the compounds and surface energy calculations with corresponding Miller indices, as discussed in the surface benchmark section. The surface energy values are in units of eV/˚A2 . | mp-id | composition | miller index | M3GNet | CHGNet | MACE | DFT | | --- | --- | --- | --- | --- | --- | --- | mp-126 mp-126 mp-126 mp-126 mp-126 mp-126 mp-1639 mp-733 mp-733 mp-733 mp-733 mp-1138 mp-1138 mp-1138 mp-1138 mp-1138 mp-19009 mp-19009 mp-19009 mp-19009 mp-19009 mp-19399 mp-19399 mp-19399 | Pt | [1, 1, 1] | 0.034337 | 0.067933 | 0.108786 | | --- | --- | --- | --- | --- | | Pt | [2, 2, 1] | 0.041738 | 0.077731 | 0.120055 | | Pt | [2, 1, 0] | 0.055588 | 0.087461 | 0.132217 | | Pt | [1, 1, 0] | 0.049405 | 0.084198 | 0.129928 | | Pt | [2, 1, 1] | 0.044232 | 0.079317 | 0.122664 | | Pt | [1, 0, 0] | 0.045627 | 0.080184 | 0.129639 | | BN | [1, 1, 0] | 0.093675 | 0.081441 | 0.127251 | | GeO2 | [1, 0, 0] | 0.073282 | 0.081616 | 0.079633 | | GeO2 | [0, 0, 1] | 0.087457 | 0.018861 | 0.080309 | | GeO2 | [1, 0, 1] | 0.066760 | 0.064233 | 0.062573 | | GeO2 | [2, 0, 1] | 0.073349 | 0.076060 | 0.083530 | | LiF | [2, 1, 0] | 0.030142 | 0.025535 | 0.028741 | | LiF | [2, 1, 1] | 0.057645 | 0.047527 | 0.056996 | | LiF | [2, 2, 1] | 0.055239 | 0.043816 | 0.052588 | | LiF | [1, 1, 0] | 0.040302 | 0.034503 | 0.040202 | | LiF | [1, 0, 0] | 0.017387 | 0.013148 | 0.015115 | | NiO | [1, 0, 0] | 0.041740 | 0.034537 | 0.078835 | | NiO | [1, 1, 0] | 0.083462 | 0.069599 | 0.112546 | | NiO | [2, 1, 0] | 0.068462 | 0.054417 | 0.102001 | | NiO | [2, 1, 1] | 0.104637 | 0.087699 | 0.126782 | | NiO | [2, 2, 1] | 0.108441 | 0.088199 | 0.129272 | | Cr2O3 | [2, 1, 1] | 0.088672 | 0.087783 | 0.103788 | | Cr2O3 | [0, 0, 1] | 0.074112 | 0.070402 | 0.085364 | | Cr2O3 | [1, 0, 0] | 0.091601 | 0.090690 | 0.099994 | 0.094540 0.111288 0.124666 0.124951 0.110808 0.129515 0.173043 0.089101 0.029946 0.079778 0.094763 0.033885 0.067411 0.058832 0.047889 0.019274 0.075694 0.146961 0.116266 0.179743 0.176103 0.137950 0.103951 0.135110 mp-19399 mp-19399 mp-19399 mp-124 mp-124 mp-124 mp-124 mp-124 mp-124 mp-841 mp-841 mp-841 mp-841 mp-841 mp-1143 mp-1143 mp-1143 mp-1143 mp-1143 mp-1143 mp-30 mp-30 mp-30 mp-30 mp-30 mp-30 mp-1894 mp-1894 mp-13 mp-13 | Cr2O3 | [2, 1, -2] | 0.100278 | 0.108988 | 0.115123 | | --- | --- | --- | --- | --- | | Cr2O3 | [2, -1, 2] | 0.100240 | 0.099631 | 0.113920 | | Cr2O3 | [2, 0, -1] | 0.083573 | 0.083330 | 0.095162 | | Ag | [2, 1, 0] | 0.023318 | 0.041564 | 0.040281 | | Ag | [2, 1, 1] | 0.019012 | 0.036769 | 0.035553 | | Ag | [1, 1, 1] | 0.014562 | 0.030217 | 0.028861 | | Ag | [1, 1, 0] | 0.021264 | 0.039141 | 0.037516 | | Ag | [1, 0, 0] | 0.019419 | 0.034914 | 0.034974 | | Ag | [2, 2, 1] | 0.018113 | 0.036006 | 0.034438 | | Li2O2 | [1, 0, 2] | 0.046675 | 0.029191 | 0.039260 | | Li2O2 | [2, 1, 2] | 0.051622 | 0.032755 | 0.033052 | | Li2O2 | [1, 0, 0] | 0.086174 | 0.063395 | 0.065709 | | Li2O2 | [2, 1, 0] | 0.059645 | 0.038821 | 0.060778 | | Li2O2 | [1, 1, 0] | 0.047108 | 0.030317 | 0.044707 | | Al2O3 | [1, 0, -2] | 0.086088 | 0.080818 | 0.103740 | | Al2O3 | [2, 1, -2] | 0.164204 | 0.146616 | 0.166867 | | Al2O3 | [2, 0, -1] | 0.106417 | 0.097795 | 0.116902 | | Al2O3 | [1, 0, 0] | 0.111818 | 0.102915 | 0.119333 | | Al2O3 | [2, 1, 2] | 0.116533 | 0.104785 | 0.130883 | | Al2O3 | [2, -1, 2] | 0.128921 | 0.124779 | 0.133044

FAQs

What problem does the supplementary material focus on?

It focuses on overcoming systematic softening in universal machine learning interatomic potentials using fine-tuning.

What does the Surface Benchmark section contain?

It presents compounds and surface energy calculations, reporting values in eV/Å² together with Miller indices across multiple models and DFT.

How does fine-tuning compare to training from scratch in the document?

A dedicated section contrasts fine-tuning with training from scratch to evaluate differences in performance and softening behavior.

Aug 3, 2026

Upload by Liam