Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Matlantis-PFP v8: Universal Machine Learning Interatomic Potential with Better Experimental Agreements via r2SCAN Functional

Chikashi Shinagawa, So Takamoto, Daiki Shintani, Yong-Bin Zhuang, Yuta Tsuboi, Katsuhiko Nishimra, Kohei Shinohara, Shigeru Iwase, Yuta Tanaka, Ju Li

Abstract

Universal Machine Learning Interatomic Potentials (uMLIPs) enable atomistic simulations and high-throughput screening at scales far beyond those accessible with density functional theory (DFT). However, most existing uMLIPs are trained on Perdew--Burke--Ernzerhof (PBE) generalized gradient approximation (GGA) data and are therefore fundamentally limited by PBE-level accuracy. In this paper, we argue that better zero-shot predictions versus experiments must be an explicit design target for uMLIPs and present PFP v8, a uMLIP available on the Matlantis service that overcomes the inherent limitations of the PBE functional by being trained to reproduce the regularized-restored strongly constrained and appropriately normed (r2SCAN) meta-GGA potential-energy surface across a wide range of chemical domains. Without requiring domain-specific fine-tuning, PFP v8 delivers systematically improved agreement with experimental data or high-accuracy references for crystals, molecules, and surfaces, outperforming PBE-based DFT calculations. Crucially, in long-time molecular dynamics simulations that are computationally impractical with DFT, PFP v8 predicts melting points with an average error of approximately 130 K, halving the error relative to PBE-trained models. These results establish that uMLIPs can move beyond the limitations of their training approximations and achieve substantially improved agreement with experiment across diverse chemical domains, further narrowing the gap between simulation and reality.

Matlantis-PFP v8: Universal Machine Learning Interatomic Potential with Better Experimental Agreements via r2SCAN Functional

Abstract

Universal Machine Learning Interatomic Potentials (uMLIPs) enable atomistic simulations and high-throughput screening at scales far beyond those accessible with density functional theory (DFT). However, most existing uMLIPs are trained on Perdew--Burke--Ernzerhof (PBE) generalized gradient approximation (GGA) data and are therefore fundamentally limited by PBE-level accuracy. In this paper, we argue that better zero-shot predictions versus experiments must be an explicit design target for uMLIPs and present PFP v8, a uMLIP available on the Matlantis service that overcomes the inherent limitations of the PBE functional by being trained to reproduce the regularized-restored strongly constrained and appropriately normed (r2SCAN) meta-GGA potential-energy surface across a wide range of chemical domains. Without requiring domain-specific fine-tuning, PFP v8 delivers systematically improved agreement with experimental data or high-accuracy references for crystals, molecules, and surfaces, outperforming PBE-based DFT calculations. Crucially, in long-time molecular dynamics simulations that are computationally impractical with DFT, PFP v8 predicts melting points with an average error of approximately 130 K, halving the error relative to PBE-trained models. These results establish that uMLIPs can move beyond the limitations of their training approximations and achieve substantially improved agreement with experiment across diverse chemical domains, further narrowing the gap between simulation and reality.
Paper Structure (14 sections, 1 equation, 4 figures, 5 tables)

This paper contains 14 sections, 1 equation, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Dataset domains and structures
  4. DFT calculation
  5. MLIP Architecture
  6. Training with multiple datasets
  7. DFT-D3 Correction
  8. Results and Discussion
  9. Validation of r$^2$SCAN-level accuracy across diverse chemical domains
  10. Crystal formation energies
  11. Molecular benchmarks: GMTKN55
  12. Surface energies
  13. Bridging simulation and reality: melting points calculation
  14. Conclusion

Figures (4)

  • Figure 1: Overview of our original datasets. Examples of structures included in the PFP-R2SCAN dataset: (a) molecule, (b) bulk, (c) disorder, and (d) slab structures, visualized by VESTA momma2011vesta. (e) Elemental coverage of our original datasets. Elements highlighted in both red and blue are included in both the PFP-R2SCAN and PFP-PBE/+U datasets (70 elements). Elements highlighted in blue are included only in the PFP-PBE/+U dataset.
  • Figure 2: Reproducibility of the experimental surface energy using (a) PFP-R2SCAN and (b) PFP-PBE. Error bars indicate experimental uncertainties.
  • Figure 3: Examples of solid-liquid interface structures, visualized by VESTAmomma2011vesta, for melting-point determination of (a) metals, (b) covalent solids, (c) oxides, and (d) ionic compounds.
  • Figure 4: Comparison of melting temperatures determined by (a) PFP-R2SCAN and (b) PFP-PBE with experimental values. The MAEs including and excluding the data points for Au and Pt are shown due to underfitting of PFP to the two materials. Details are summarized in Table \ref{['tab:exp_melting_points']} and Table \ref{['tab:calc_melting_points']}.