Technical Note on Matlantis PFP v9: Extending the Element and Domain Coverage of the r2SCAN Training Data
Matlantis PFP v9 extends the r2SCAN-trained universal machine learning potential to cover 96 elements (including lanthanides and actinides) and new domains like clusters and adsorption, achieving experimental-level accuracy for formation energies and DFT-comparable adsorption errors while significantly outperforming SevenNet-Omni in inference speed and maximum system size.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer
In the microscopic world of atoms, scientists have long relied on a powerful but heavy tool called density functional theory to predict how matter behaves. This method calculates the interactions between every single atom in a material with extreme precision, essentially solving the quantum mechanical equations that govern their existence. While this approach yields results that match real-world experiments, it is so computationally demanding that it can only handle systems containing a few thousand atoms for a brief moment in time. To study larger structures, like a grain of metal or a drop of water, or to watch processes unfold over longer periods, researchers needed a faster way to get nearly the same level of accuracy. This need gave rise to machine learning interatomic potentials, which are computer models trained on the data from those heavy calculations. Once trained, these models can predict how atoms will move and interact at a speed thousands of times faster than the original method, allowing scientists to simulate systems with millions of atoms and explore vast new territories of material science.
For years, these machine learning models were often specialized, built to handle only specific types of materials like organic molecules or simple crystals. A significant leap occurred with the development of Matlantis PFP, a universal model designed to treat a wide variety of domains—from solid crystals to individual molecules and surface layers—with a single, unified system. The latest version of this model, known as PFP v9, represents a major expansion of its capabilities. Researchers at Preferred Networks and Matlantis Corporation have successfully taught this model to understand a much broader range of the periodic table and more complex physical arrangements. By feeding the model a new, massive dataset calculated with a highly accurate method called r2SCAN, they have enabled it to simulate materials containing rare-earth elements and heavy actinides, which were previously too difficult to model with this specific approach. They also added data on how molecules stick to surfaces and how atoms cluster together, filling in gaps that limited the model's previous usefulness.
The core of this advancement lies in the training data. The researchers generated a new set of reference calculations using the r2SCAN functional, a sophisticated mathematical approach that agrees very well with experimental results. In the previous version of the model, this dataset covered only 70 elements and missed entire families of atoms, such as the lanthanides, which are crucial for technologies like magnets and batteries. The new work extended this coverage to include 96 elements in total, adding the lanthanides from praseodymium to lutetium and heavy elements from polonium to curium. To handle these complex atoms, the team made a specific technical choice regarding how to treat their inner electrons. For the lanthanides, they used "hard" potentials that allow the inner f-electrons to participate in bonding, rather than keeping them locked away in a frozen core. This decision was necessary to accurately describe how these atoms behave in different chemical environments, a nuance that had been left unresolved in earlier versions.
With the model now trained on this expanded and more diverse dataset, the researchers tested its accuracy against real-world measurements. They looked at the formation energies of crystals containing the newly supported elements, which is a measure of how much energy is released or absorbed when a material forms from its constituent parts. The results showed that the model's predictions for these new, complex materials were just as accurate as its predictions for the older, well-known elements. The difference between the model's calculations and experimental values was small, averaging about 93 millielectronvolts per atom, a figure nearly identical to the error rate for the previously supported elements. This indicates that the model has successfully learned the physics of these difficult-to-handle atoms without losing its precision.
The team also tested the model's ability to predict adsorption, which is the process where molecules stick to a solid surface. This is a critical interaction in fields like catalysis, where surfaces are used to speed up chemical reactions. They compared the model's predictions for how strongly various molecules bind to transition-metal surfaces against a database of experimental measurements. The model performed with an overall error comparable to the standard density functional theory method used as a benchmark. However, the researchers noted a specific tendency: the model slightly overestimated the strength of the bond for molecules that form strong chemical bonds with the surface, while it underestimated the bond strength for those held together by weaker forces. Despite this, the inclusion of a dispersion correction, which accounts for weak attractive forces between atoms, brought the predictions for the weaker interactions much closer to experimental reality.
Beyond accuracy, the practical speed of the model is a defining feature of this work. The researchers compared PFP v9 against another public model trained on similar data, known as SevenNet-Omni, running both on the same type of high-performance graphics processor. The results were striking. The new model could handle systems containing five to ten times as many atoms as the competitor before running out of memory. Furthermore, when calculating the forces that drive atomic motion, the new model was three to four times faster. For a system of liquid water, for instance, the model could simulate a cluster of nearly 118,000 atoms, whereas the other model could only manage about 12,000 before hitting a memory limit. This leap in efficiency means that scientists can now simulate much larger and more complex systems, such as rare-earth alloys or actinide compounds, with a speed that was previously impossible for this level of accuracy.
The authors conclude that this expansion makes the model a versatile tool for a wide array of scientific challenges, from designing new materials for energy storage to understanding the behavior of nuclear fuels. However, they are clear about the boundaries of what the model can currently do. While it excels at describing atoms in their standard states and on clean surfaces, it still struggles with certain complex scenarios, such as materials with strong magnetic correlations or systems involving long-range electrical charges. The model does not yet account for the effects of electric fields or the behavior of charged defects in a way that would allow it to simulate electrolytes or electrified interfaces. These remain areas for future development. For now, the work stands as a significant step forward, providing a reliable and fast way to explore the properties of a vast swath of the periodic table that was previously out of reach for universal machine learning models.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.