Training datasets for Matlantis’s core AI technology are now developed using r²SCAN, doubling accuracy in atomistic simulations compared to previous version
CAMBRIDGE, Mass., July 16, 2025 — , the U.S. base for the materials discovery division of Japan’s leading AI company Preferred Networks, Inc. (PFN), today announced a significant update to its Matlantis™ universal atomistic simulator, along with the opening of its office in Cambridge, Massachusetts. This move aims to accelerate the adoption of AI-driven materials research across North America. The update introduces Version 8 of PFN’s proprietary AI technology, PFP (Preferred Potential), which empowers researchers across various industries to expedite discovery, enhance predictive performance, and explore new frontiers in materials science with unprecedented levels of simulation accuracy.
PFP Version 8 represents a major milestone, as it is the first universal machine learning interatomic potential (MLIP) to be trained using datasets developed with a new method called r2SCAN (restored-regularized strongly constrained and appropriately normed) functional. Previous PFP versions, up to 7, relied on datasets generated with the PBE (Perdew-Burke-Ernzerhof) functional, a method also widely adopted by MLIPs other than PFP. However, PBE is known to have certain limitations regarding simulation accuracy—referring to how closely computer-based simulations of materials’ behavior align with real-world experimental results.
The introduction of the r2SCAN method is the culmination of PFN’s continuous efforts over the past few years to overcome the accuracy limitations of the PBE-based approach. Developing training datasets with the r2SCAN method is more computationally intensive, requiring three to five times the computing time compared to the PBE method. Nevertheless, because PFP Version 8 is now trained with datasets built using both r2SCAN and PBE, Matlantis users can achieve up to double the simulation accuracy within the same timeframe as the previous version.
“This update represents a significant breakthrough,” stated , CEO of Matlantis. “In 2021, we were the first globally to launch a commercial simulator utilizing a universal MLIP, and now our simulator, Matlantis, is the first worldwide to incorporate r2SCAN, which ensures high simulation accuracy. We believe this will further pave the way for the era of computer-based materials discovery. We will continue to support researchers in North America and the rest of the world in discovering innovative and sustainable new materials.”
Matlantis, a joint investment by PFN, Japan’s largest energy company ENEOS, and Mitsubishi Corporation, has been adopted by over 100 industrial and academic leaders globally since its launch in July 2021. Today, Matlantis is among the first commercially available AI-powered platforms specifically designed for industrial-scale atomistic simulation—offering a single MLIP that covers 96 elements (from hydrogen to curium) and provides DFT (density functional theory)-level accuracy up to 20 million times faster.
Matlantis empowers research teams to:
- Perform simulations from the first day of use:
Matlantis is provided as a cloud-based software-as-a-service (SaaS). Users can access it via a browser and begin searching for new materials from their first day of use. Because Matlantis’s machine learning interatomic potential (MLIP) has already been trained with extensive datasets, users can immediately focus on material discovery without spending time building machine learning models.
- Search a wide variety of undiscovered materials
As a universal atomistic simulator, Matlantis covers a broad range of materials for batteries, semiconductors, catalysts, and more, without requiring changes to AI models based on material types.
- Accelerate materials discovery
With Matlantis, researchers can complete simulations in just a few hours that would otherwise necessitate years of conventional DFT calculations. This speed-up transforms iterative design in materials discovery, reorienting the R&D process so that computational insights lead experiments, rather than merely validating them afterward.
- Achieve higher simulation accuracy than ever before
With the new training datasets built using the r2SCAN method, Matlantis can simulate material properties with greater accuracy than common MLIPs within the same timeframe, further narrowing the gap between simulations and experiments.
“With PFP 8.0, we finally have a universal machine learning interatomic potential that maintains the best DFT-level fidelity while spanning most of the periodic table,” said Matlantis Technical Advisor Prof. Ju Li, Ph.D., widely recognized for his work on atomistic modeling and materials research. “That combination of accuracy and speed allows engineers to generate phase diagrams or screen multi-component systems in hours or several days, instead of weeks or months—work that directly informs alloy design, battery materials, and other high-value applications. Establishing a U.S. office means we can collaborate even more closely with industrial and academic partners here, shorten feedback loops, and bring new Matlantis capabilities to market faster.”
Dr. Katsushisa Yoshida, Director and Deputy Head of the Research Center for Computational Science and Informatics at Resonac, remarked: “We are excited to hear about the major update to Matlantis and the opening of their new U.S. office. We eagerly anticipate how the evolution of this platform will further accelerate our own materials development.”
PFP 8.0 was developed using PFN’s supercomputer and 2.0 and 3.0, provided by Japan’s National Institute of Advanced Industrial Science and Technology (AIST) and AIST Solutions Co., Ltd. The use of ABCI 3.0 is supported by the ABCI 3.0 Development Acceleration Program.
About Matlantis
Matlantis, jointly developed by PFN and ENEOS, is a universal atomistic simulator that supports large-scale material discovery by reproducing the behavior of new materials at an atomic level on the computer. PFN and ENEOS have integrated a deep learning model into a conventional physical simulator to increase simulation speed by tens of thousands of times and to support a wide variety of materials. Matlantis was launched in July 2021 as a cloud-based software-as-a-service by Matlantis Inc. (formerly named Preferred Computational Chemistry), a company jointly invested by PFN, ENEOS, and Mitsubishi Corporation. Matlantis is used by over 100 companies and organizations for discovering various materials including catalysts, batteries, semiconductors, alloys, lubricants, ceramics, and chemicals.
For more information, please visit:
Media Contact:
Janabeth Ward
Scratch Marketing + Media for Matlantis
