Comparison with Open Source MLIP
Modeling key phenomena in real-world materials development—such as reactions and diffusion at interfaces—can require more than modeling with a small number of atoms or short molecular dynamics simulations. Therefore, when evaluating the performance of an MLIP, the number of atoms it can simulate and its computational speed are critical performance metrics.
Comparison with OSS-MLIPs (Pt fcc)
Here, we present a comparison of PFP v9 and OSS MLIPs in terms of the maximum number of atoms that can be simulated and computational speed. PFP was executed on the Matlantis platform, while the other MLIPs were run using an NVIDIA A100 GPU (80 GB memory) with the ASE Calculator. The number of atoms was increased by supercell expansion of fcc Pt, and calculation speed was measured as the average time of 10 force evaluations.
As a result, PFP showed the fastest computation time for the same number of atoms, with a more gradual increase in computation time as the number of atoms grows. The figure and table below show the comparison (the figure displays only results with 50,000 atoms or fewer and 5 seconds or less for visibility).
The maximum number of atoms for each model is listed below. Note that the maximum number of atoms for PFP is set by the system-size limit accepted by Matlantis, not by GPU memory. UMA-S can handle larger systems than PFP, but at about 30,000 atoms its speed is roughly one sixth that of PFP, respectively.
Table 1. Maximum number of atoms supported by each model
| Model (environment) | Max. atoms |
|---|---|
| PFP v9 (Matlantis) | 42,875 |
| SevenNet-Omni (A100) | 19,683 |
| UMA-S (A100) | 97,336 |
| MACE-MPA-0 (A100) | 13,824 |
| MatterSim-v1-5M (A100) | 10,648 |
| CHGNet (A100) | 4,913 |
| eqV2-31M OMat MPtrj-sAlex (A100) | 1,728 |
| Orb-v3 cons-inf-omat (A100) | 10,648 |
Detailed comparison with SevenNet-Omni
We compared SevenNet-Omni (matpes_r2scan task), a public uMLIP trained on r²SCAN data, with the r²SCAN calculation mode of PFP (hereafter PFP-R2SCAN) in detail on three systems: fcc Pt, rocksalt LiH (lattice constant 4.083 Å), and liquid water (a cubic cell of eight molecules). The calculation conditions are the same as above.
The table below summarizes the time required for force inference at the largest size SevenNet-Omni could handle for each system (given under the system name). For all three systems, the maximum number of atoms PFP-R2SCAN can handle is about twice that of SevenNet-Omni, and force evaluation at the same size is also faster; the largest difference is for liquid water, where PFP-R2SCAN is about 2.4 times faster at 12,288 atoms. Systems with a higher number density also tend to take longer to process.
Table 2. Detailed speed benchmark results for PFP and SevenNet-Omni
| System | Max. atoms | Time per force evaluation (s/step) | ||
|---|---|---|---|---|
| PFP-R2SCAN (Matlantis) | SevenNet-Omni (A100) | PFP-R2SCAN (Matlantis) | SevenNet-Omni (A100) | |
| Pt fcc (at 19,683 atoms) | 44,100 | 19,683 | 1.29 | 1.99 |
| LiH rocksalt (at 11,664 atoms) | 24,334 | 11,664 | 0.89 | 1.90 |
| Liquid water (at 12,288 atoms) | 29,040 | 12,288 | 0.70 | 1.68 |
These results suggest that, compared with OSS-MLIPs, PFP enables simulations of larger systems as well as longer simulations of condensed-phase systems. This capability is expected to accelerate both the understanding of complex materials behavior and the accurate prediction of physical properties.
-
URL
Copied
