image: The 40,000 entries of the 0123D test set.
Credit: Benedini et al, AI for Science (2025).
Bochum, Germany, October 29, 2025, Researchers from Research Center for Future Energy Materials and Systems at the Ruhr University Bochum, Software for Chemistry & Materials BV, and Vrije Universiteit Amsterdam have demonstrated that modern universal machine learning interatomic potentials (uMLIPs) can now accurately describe systems ranging from single molecules to bulk solids, representing a significant leap forward for uMLIPs in materials science. The study introduces the 0123D dataset, comprising 40,000 diverse structures specifically designed to benchmark model performance across all dimensionalities.
Researchers from Ruhr University Bochum and partners have developed a comprehensive benchmark evaluating how well artificial intelligence models simulate materials across different dimensions, from individual molecules to bulk crystals. Traditional quantum mechanical calculations are highly accurate but computationally expensive. Machine learning interatomic potentials offer comparable accuracy at a fraction of the cost, potentially revolutionizing materials discovery for batteries, catalysts, and other applications. The study tested eleven state-of-the-art machine learning models using a new dataset of 40,000 structures spanning molecules, nanowires, atomic layers, and bulk solids. Results show that while most models excel at simulating three-dimensional materials, accuracy decreases significantly for lower-dimensional structures like molecules and nanowires. The equivariant Smooth Energy Network (eSEN) emerged as the standout performer, maintaining remarkable accuracy across all dimensionalities with atomic position errors of just 0.01–0.02 Å and energy errors below 10 meV/atom, values that approach the precision of quantum mechanical calculations. These results demonstrate that the best AI models can already serve as direct replacements for expensive quantum calculations in many applications. The exceptional performance of eSEN across all dimensions opens new possibilities for modeling complex interfaces and catalytic reactions. The performance degradation in most models stems from training data bias toward three-dimensional structures. The findings suggest that more diverse training datasets could significantly improve model transferability. The 0123D benchmark dataset is freely available at https://alexandria.icams.rub.de/ to enable continued model development and evaluation. Comprehensive benchmark results are provided at https://github.com/GiulioIlBen/0123D-uMLIPs-benchmark.
Contact: Prof. Dr. Silvana Botti, Ruhr University Bochum, silvana.botti@rub.de
Reference: Giulio Benedini, Antoine Loew, Matti Hellström, Silvana Botti and Miguel Marques. Universal Machine Learning Potentials for Systems with Reduced Dimensionality[J]. AI For Science, 2025, 1(2). DOI: 10.1088/3050-287X/ae1208