PFP Version History

Introducing PFP v7.0.0

PFP v7.0.0 was officially released in September 2024, expanding the number of supported elements from 72 to 96.

These 96 elements cover the entire set of naturally occurring elements. Combined with its high prediction accuracy and the ability to handle large-scale simulations of up to 44,000 atoms, PFP now enables virtually all molecular dynamics (MD) simulations commonly required in materials development.

Our Journey So Far

Here we look back at the evolutionary path that PFP has taken.

Release Date	2021.07	2022.08	2023.10	2024.04	2024.09
PFP Version	v1	v3	v5	v6	v7	Cumulative Progress
Key Topics	D3 Correction (v1.1)	Expanded range of supported elements PFVM applied	Improved molecular interactions	New Architecture	Expanded range of supported elements
Corresponding element type	55	72	72	72	96	+41 elements
Number of atoms that can be calculated*	3,000	10,000	19,000	44,000	44,000	Approximately 15 times
Number of training data	10 million	22 million	42 million	42 million	59 million	Approximately 6 times

PFP v1.0.0

PFP v1.0.0, released in July 2021, marked a breakthrough in atomistic simulation for materials development. It introduced a truly universal machine learning potential—a single model capable of handling a wide range of material systems—challenging the conventional approach that required developing separate interatomic potentials for each material. At the time of release, it supported 55 elements and systems of up to 3,000 atoms. This initial version laid the foundation for continuous improvements in performance. In v1.1, support for DFT-D3 dispersion corrections was added, enabling the model to account for long-range interactions.

PFP v3.0.0

PFP v3.0.0 was released in August 2022. By expanding the training dataset to 22 million structures, the number of supported elements was increased to 72. Additionally, memory efficiency was significantly improved through the introduction of PFVN, a custom optimization technology developed by Preferred Networks, enabling simulations of systems with up to 10,000 atoms.

PFP v5.0.0

PFP v5.0.0 was released in October 2023. The training dataset was expanded to 42 million structures, which were used to further improve the model. As a result, the model's ability to reproduce intermolecular interactions was significantly improved.

PFP v6.0.0

PFP v6.0.0 was released in April 2024. A redesigned architecture significantly increased the maximum number of atoms that can be simulated—from previous limits to up to 44,000 atoms. This enabled large-scale simulations that were previously out of reach.

The Future of PFP

PFP will continue to be updated. We are considering three goals: 1) improving performance, 2) increasing speed, and 3) providing technology to broaden the scope of research. Here, we will introduce our efforts toward 1) improving performance.
Conventional PFP has used the results of DFT calculations using the PBE functional for much of the training data. However, customer accuracy requirements have reached a point where the discrepancy between PBE calculations and experimental values is no longer acceptable.
Therefore, we are considering adding DFT calculation results obtained using the r2SCAN functional, which is more accurate than PBE, to the learning data of the next PFP. DFT calculations using r2SCAN require an enormous amount of calculation time, but by having PFP learn these highly accurate results, customers will be able to run simulations with r2SCAN-level accuracy while maintaining the same calculation costs as before.
For more details on our future development roadmap, please contact us.