Support for 96 Elements and Improved Robustness of Universal Neural Network Potential PFP

The concept of a universal interatomic potential that can be applied to various systems without limiting the domain has attracted attention, and is being developed by many groups as a fundamental model of atomic systems. The neural network potential PFP being developed by the authors is characterized by its universality in handling arbitrary combinations of elements in a single model, rather than performing fine tuning for each system. Both dataset expansion and architectural improvements are ongoing, with the aim of achieving both greater universality and accuracy.
In terms of universality, the number of elements supported by PFP has been expanded sequentially with the expansion of the dataset, starting with the initial 45 elements in 2022 and then expanding to 55 and 72 elements, and the PFP currently under development is being expanded to 96 elements, up to the transuranium element Cm (curium), which will cover all elements stably present on earth. In terms of the elements, we have reached an area that can be regarded as truly universal.
Robustness is also important in simulations to ensure that the state does not go to extrapolated regions. In particular, in exploratory tasks such as crystal structure prediction, energetically unstable structures such as those with multiple atoms positioned closer together may be generated, and the ability of the stable inference in such a situation is required. Architectures based on graph neural networks tend to be unstable for compressed structures because the graph representation is excessively dense, and PFP has tackled this problem both in terms of the architecture and the dataset, and has significantly extended the range of stable inference in PFP v6.0.0.