AI-Driven Expansion and Application of the Alexandria Database

AI-Driven Expansion and Application of the Alexandria Database Théo Cavignac,1 Jonathan Schmidt,2 Pierre-Paul De Breuck,1 Antoine Loew,1 Tiago F. T. Cerqueira,3 Hai-Chen Wang,1 Anton Bochkarev,4 Yury Lysogorskiy,4 Aldo H. Romero,5 Ralf Drautz,4 Silvana Botti,1, ∗and Miguel A. L. Marques1, † 1Research Center Future Energy Materials and Systems of the University Alliance Ruhr and ICAMS, Ruhr University Bochum, Universitätsstraße 150, D-44801 Bochum, Germany 2Department of Materials, ETH Zürich, Zürich, CH-8093, Switzerland 3CFisUC, Department of Physics, University of Coimbra, Rua Larga, 3004-516 Coimbra, Portugal 4ICAMS, Ruhr-Universität Bochum, Universitätstrasse 150, 44801 Bochum, Germany and ACEworks GmbH, Hagen-Hof-Weg 1, 44797 Bochum, Germany 5Department of Physics, West Virginia University, Morgantown, WV 26506, USA (Dated: December 11, 2025) We present a novel multi-stage workflow for computational materials discovery that achieves a 99% success rate in identifying compounds within 100 meV/atom of thermodynamic stability, with a threefold improvement over previous approaches. By combining the Matra-Genoa generative model, Orb-v2 universal machine learning interatomic potential, and ALIGNN graph neural network for energy prediction, we generated 119 million candidate structures and added 1.3 million DFT-validated compounds to the Alexandria database, including 74 thousand new stable materials. The expanded Alexandria database now contains 5.8 million structures with 175 thousand compounds on the convex hull. Predicted structural disorder rates (37–43%) match experimental databases, unlike other recent AI-generated datasets. Analysis reveals fundamental patterns in space group distributions, coordination environments, and phase stability networks, including sub-linear scaling of convex hull connectivity. We release the complete dataset, including sAlex25 with 14 million out-of-equilibrium structures containing forces and stresses for training universal force fields. We demonstrate that fine-tuning a GRACE model on this data improves benchmark accuracy. All data, models, and workflows are freely available under Creative Commons licenses. I. INTRODUCTION Recent advances in availability of large materials databases serves as evidence of the success of high-throughput (HT) materials discovery [1–8] and provide reservoirs of hypothetical compounds that are thermodynamically stable or close to stability. Despite these advances, sampling the vast combinatorial space of possible materials remains computationally intensive and unfeasible by brute force. Fortunately, the development of machine learning (ML) models has greatly accelerated this process [9–13]. Conversely, we have also witnessed a dramatic increase of data available to train new models [3, 7, 8, 14]. Such a positive feedback loop has significantly accelerated the material discovery for technological applications ranging from energy storage to catalysis and electronics. The Alexandria database represents one of the largest collections of ab initio calculations and is the largest open database for thermodynamically stable materials. Currently, it encompasses approximately 5.8 million structures calculated using density functional theory (DFT), of which 175 thousand structures lie on the convex hull. The previous iteration of this dataset has been ∗silvana.botti@rub.de † miguel.marques@rub.de extensively utilized by the scientific community. For instance, all universal force-field models within the top five [10, 14–18] of the Matbench Discovery ranking [19] were trained using Alexandria data. Some of these force fields have advanced to the point where they can reliably predict structural, vibrational, and thermal properties [20], defect energies [21, 22], and infrared spectra [23], thereby unlocking new avenues for exploring physical phenomena. Beyond training machine-learning force fields, Alexandria has also been employed in the development of generative models [24]. In addition to serving as a training resource, this extensive database provides a valuable foundation for data-driven design of functional materials. For instance, subsets of Alexandria have been used to identify novel dielectric semiconductors [25], novel two dimensional materials [26] and hard magnets [27]. Furthermore, state-of-the-art machine- learning–assisted high-throughput design frameworks for conventional high-Tc superconductors leverage Alexandria as their primary search space [28–30]. A major challenge in the growth of Alexandria is to generate novel structures that are potentially located near the convex hull with high success rate. During standard prototype-based high-throughput studies the success rate is of the order of 0.1% [31]. To increase the success rate of standard HT search, Wang et al. [32] used data-driven chemical similarity measures to guide systematic substitution, i.e. replacing elements in stable or near-stable compounds with similar elements. This arXiv:251

…(Full text truncated)…

This content is AI-processed based on ArXiv data.