China has unveiled the latest version of its large-scale artificial intelligence model, after it secured the top spot on Matbench Discovery, a leading international benchmark for materials discovery.
The atomic AI model, DPA4, achieved the highest overall performance on the platform, which is widely used by researchers to evaluate how effectively AI systems can discover and predict new materials. The achievement highlights China’s growing efforts to use artificial intelligence to accelerate scientific research and materials development.
DPA4 was developed by a joint team from the AI for Science Institute (AISI) in Beijing, DP Technology, Peking University, and the Institute of Applied Physics and Computational Mathematics.
The model belongs to the Deep Potential Atomic (DPA) family, an AI system designed to simulate atomic interactions and material behavior at the atomic level. Researchers say the latest version delivers high accuracy while requiring far less computing power, making advanced materials research faster and more efficient.
The model has undergone four generations of upgrades over the past four years, with DPA4 representing its latest advancement.
READ ALSO: Meet ‘Heaven Sword’: The 84-Meter Giant Crowned East Asia’s Tallest Tree in Taiwan’s Mountains
According to the research team, the model’s main advantage is its ability to maintain high accuracy while operating at a much lower computational cost.
Li Tiancheng, a doctoral student at Peking University and a core member of the project, said DPA4 improves training efficiency by about ten times compared with DPA3 at the same accuracy level. This allows researchers to process large datasets faster and more efficiently.
Faster Simulations for Research
DPA4 can simulate changes in energy and force interactions between atoms in a variety of materials. Its current capabilities cover inorganic crystals and small organic molecules, two important categories in scientific and industrial research. These simulations help scientists understand how materials behave under different conditions.
The model can support research in several major sectors. These include battery materials, catalysts, semiconductors, and pharmaceutical molecules. Faster and more accurate simulations can reduce development time and help researchers identify promising materials earlier in the discovery process.
Materials discovery has traditionally required extensive laboratory testing and high-performance computing resources. AI models such as DPA4 help narrow down the most promising candidates before physical testing begins. This approach saves both time and research costs while improving efficiency.
READ ALSO: Arizona Scientists Uncover the Wanderlust of America’s Iconic Bald Eagles
Open-Source Research Expansion
The research team has already released an early-access version of DPA4 through open-source communities. Scientists and developers can test the model and explore its capabilities before the official public release. The developers have confirmed that a full open-source version will be launched at a later stage.
Open access is expected to encourage wider collaboration among universities, research institutes, and industry partners. By making advanced simulation tools more accessible, researchers can accelerate innovation across multiple scientific fields. This also helps smaller organizations gain access to technology that was once available only to large research centers.
The rise of AI-powered scientific tools is reshaping how new materials are discovered and developed. As DPA4 continues to evolve, it is expected to support larger and more frequent atomic simulations, helping researchers move from traditional trial-and-error methods toward faster, data-driven scientific discovery.
