What is the AI model "Matlantis PFP"?

Matlantis PFP, a unique Universal Machine Learning Interatomic Potential developed ahead of the rest of the world.

One highly accurate method for predicting atomic behavior involves calculations using density functional theory (DFT), but this method has been hampered by its enormous computational cost. As a solution, machine learning interatomic potential (MLIP*), which uses DFT calculation results to train AI, has been studied. However, conventionally, models specialized for specific substances and conditions have been dominant, and developing a "general-purpose model" that can handle various materials has been extremely difficult.
We have been tackling this challenging problem since the early days of deep learning, and have developed "Matlantis PFP," a unique AI model with world-class accuracy and universality. The underlying technology of PFP has also been published in the international scientific journal "Nature Communications."

Published paper: Towards universal neural network potential for material discovery applicable to arbitrary combination of 45 elements (Nature Communications, 2022)

*Machine learning interatomic potentials are also referred to as Machine Learning Potential (MLP) or Neural Network Potential (NNP), in addition to the abbreviation MLIP. While MLP and NNP carry slightly different nuances, they are generally used interchangeably to refer to machine learning interatomic potentials. In this site, we use the term MLIP, which is becoming the relatively common standard in the field.

DFT-Level Accuracy at 70 Million Times the Speed

Matlantis PFP delivers DFT-level accuracy approximately 70 million times faster for a 3,000-atom system. Simulations that once took months on supercomputers can now be completed in minutes.
It also enables large-scale systems of up to 44,000 atoms* , allowing accurate simulations of complex interfaces and defect structures that were previously impractical. By breaking through the limits of accuracy, speed, and scale, PFP is transforming the landscape of materials development.

*Maximum number of Pt(fcc) atoms that can be inferred

Supports 96 Elements, Covering the Entire Range of Naturally Occurring Elements

PFP delivers exceptional versatility through a single AI model. The latest version supports 96 elements, covering all elements found in nature. This eliminates the need for users to switch models or consider restrictions based on element types.
Moreover, its applicability goes far beyond bulk crystals. PFP can seamlessly and accurately simulate a wide variety of atomic structures and states—including catalytic reactions at surfaces and interfaces, ion diffusion in battery materials, decomposition of organic molecules, and adsorption in MOF composite materials—all within a single, unified model.

Architecture and Training Dataset

Matlantis PFP's performance is supported by an advanced AI architecture and large-scale training data. The basic structure of the model employs a "graph neural network (GNN)" that recognizes atoms and their bonds as a graph. This allows it to flexibly learn the local state and interactions of each individual atom, enabling highly accurate predictions even for unknown structures.
The training data used consists of DFT calculations of a wide variety of structures performed on Preferred Networks' computing infrastructure. Not only stable structures, but also unstable and unrealistic structures were included, and PFP v9 used training data with over 63 million structures. This overwhelming "quality and quantity of data" is the source of Matlantis PFP's unprecedented universality.

PFVM: The Inference Engine That Accelerates Simulations

Matlantis's overwhelming computing speed is achieved through a combination of two factors: the superior AI model called Matlantis PFP and the execution software that maximizes its capabilities.
At its core is the proprietary inference engine "PFVM." PFVM combines the roles of a compiler that optimizes Matlantis PFP calculations using optimization techniques such as "kernel fusion," and a runtime that executes these calculations quickly and efficiently on hardware.
In this way, the capabilities of the high-performance AI model Matlantis PFP are maximized by the specially developed inference engine PFVM, resulting in a simulation environment that is unmatched by others.