Generalized global neural network potential covering the periodic table

One universal potential for all-purpose atomic simulations has been pursued for decades, but has faced extreme challenges in both reaching high representation capability of the model and constructing comprehensive potential energy surface (PES) data across the periodic table.

Recently, a research team led by Professor Zhi-Pan Liu at Fudan University achieved a key breakthrough by developing a generalized global neural network potential (GG-NN) that spans the periodic table, enabling accurate and efficient predictions for diverse atomic systems. “As a truly practical universal potential, computational efficiency and the quality of the dataset are critical, and these are central to our GG-NN.” Professor Zhi-Pan emphasized.

High-order Pair-reduced Neural Network

To explore the global PES across diverse configurations, the team newly developed a High-order Pair-reduced Neural Network (HPNN) architecture, designed to achieve high spatial discrimination between structures while maintaining high-speed PES computation.

HPNN follows the general scheme of GPU-based graph neural networks but introduces a new interaction layer to incorporate neighboring information at low computation cost. First, HPNN temporarily compresses the dimension of atomic features before the pair interaction and restores them afterward. Second, HPNN does not utilize message passing via the CG tensor product, but incorporates different orders of spherical harmonics up to l_max=6 via concatenation, effectively balancing efficiency and accuracy.

Global dataset

The training sets of released universal foundation potential models are typically constructed by combining large-scale open-source datasets such as MPtrj, OMat24, and OC20. However, it may suffer from inconsistencies across different functionals and lack coverage of highly distorted or rare configurations.

Benefiting from decades of accumulated data from the LASP project and the use of self-learning stochastic surface walking (SSW) with element replacement, Liu’s team constructed a global PES dataset of 5.84 million configurations covering 83 elements (August 2025), all computed under a unified density functional theory setup. It should be mentioned that the unbiased SSW sampling strategy ensures systematic inclusion of high-energy and strongly distorted structures, thereby enhancing the robustness of the GGNN.

Benchmark

The resulting GG-NN potential exhibits good accuracy, namely, the root-mean-square errors of 7.348 meV/atom for energy, 0.163 eV/Å for atomic forces on the global training dataset. Across representative benchmarks—including long-range interactions, reaction barriers, materials discovery, and large-scale molecular simulations—GG-NN shows clear advantages in both computational efficiency and predictive accuracy, outperforming traditional CPU-based potentials and mainstream GPU models.

These advances mark an important step toward practical universal PES modeling. The development of GG-NN provides a pathway to accelerating molecular and materials discovery through efficient, high-resolution exploration of global potential energy surfaces across the periodic table.