Atomic simulation serves as an indispensable microscope for modern science, bridging the gap between theoretical predictions and experimental results. However, the field has historically been constrained not only by the trade-off between computational cost and accuracy but also by the difficulty of translating human-readable chemical language (e.g., chemical formulas) into 3D atomic structures. Furthermore, the lack of unified workflows and intuitive interfaces often prevents non-experts from utilizing advanced tools. Although Machine Learning Potentials (MLPs) offer a solution to the accuracy-efficiency dilemma, their potential is often limited by these practical barriers and a lack of generality across chemical elements.
Herein, this study reports the implementation of LASPAI, a user-friendly, web-based platform designed to democratize atomic simulation. By integrating the Generalized Global Neural Network (GGNN) potential trained on over 6.4 million structures covering 89 elements with high-order pair-reduced neural network with Edge and Time information (HPNN-ET) diffusion generative models, LASPAI enables rapid Potential Energy Surface (PES) exploration and automated 3D structure generation from simple inputs. The platform offers a streamlined, task-oriented workflow capable of handling complex scenarios, ranging from solid structure prediction and interface identification to reaction pathway simulations. Validated through cases such as TiO₂ polymorph prediction, pyridine adsorption on γ-Al₂O₃ surface, liquid phase separation and reaction pathway prediction, LASPAI drastically reduces simulation time while maintaining high accuracy, empowering researchers to design new materials and reactions efficiently.
This work entitled “LASPAI: AI-powered Platform for the Future Atomic Simulation” was published on Acta Physico-Chimica Sinica (published on December 31, 2025).
DOI:10.1016/j.actphy.2025.100235