META-LEARN · Meta-Learned Machine-Learning Interatomic Potentials for Ab initio Engineering of Chemical and Microstructural Complexity

Materials engineers have dreamed for decades of optimizing materials starting from the quantum mechanical laws of nature. Mastering the inherent chemical and microstructural complexity promises access to outstanding properties of structural and functional materials. Machine-learning interatomic potentials (MLIPs) ignited hope by offering quantum-mechanical accuracy for systems with many atoms. However, the full capacity of MLIPs remains untapped due to the complexity of the MLIP construction process and the required simulations.META-LEARN will cut the MLIP Gordian knot and raise MLIP construction to the next level. Our vision is a meta-learning framework that unleashes the full strength of MLIPs for large-scale simulations to a broad community. We will meta-learn the optimal MLIP-construction processes by acquiring and exploiting domain-expert knowledge from various branches of advanced ab initio and large-scale simulations.META-LEARN leverages a comprehensive pool of algorithms to ensure an optimal, task-oriented accuracy-efficiency trade-off. MLIPs that account for magnetic and electronic excitations will broaden the materials spectrum. MLIP-based sampling will boost the efficiency of thermodynamic predictions. Deciphering the microstructure genome will provide the optimal training for large-scale defects. Sustainability challenges will be tackled at the limit of chemical and microstructural complexity: Multicomponent H-storage and coating materials.META-LEARN will encode the knowledge of the MLIP construction and make it openly available via the MLIP-COPILOT, a knowledge-graph-based artificial intelligence tool. The MLIP-COPILOT will provide the optimal combinations of MLIP algorithms, hyperparameters, training datasets, and training sequences for different materials and simulation tasks. The MLIP-COPILOT will remain flexibly extensible for the community beyond the project’s scope, allowing for the addition of new types of simulations, materials, and MLIPs.