Artificial Intelligence for Intermetallic Materials

点击翻译按钮获取中文摘要

Materials science contributes to facing the challenges of our society and particularly the climate change by delivering highly optimized structural and functional materials. The relevant material classes range from structural materials like superalloys and light-weight steels for increased efficiency in aviation, automobiles and power generation to functional materials like solar cell and battery materials for electric-power generation and storage respectively. The key challenge for materials design and optimization is the prediction of the thermodynamic stability of a compound, i.e. the prediction of the energetically stable crystal structure from only its chemical composition. The enormous numerical effort involved in this prediction sets a fundamental limit on the computational exploration of materials. While density-functional theory calculation (DFT) revolutionized materials science and provides highly reliable predictions of thermodynamic stability, it is still too computationally expensive to explore the chemical and structural complexity that is needed for many technologically relevant materials. With the advent of data-driven scientific discovery, we are witnessing the change to the fourth paradigm in science. The enormous potential of data exploitation with artificial intelligence (AI) will change many fields of our society ranging from health to environment and technology. One of the fields that are changing gears with AI already now is materials science. In this project, we apply modern AI techniques to the prediction of the thermodynamic stability of intermetallic phases that play a central role in superalloys and light-weight steels. As a novel and unique approach, we will develop and apply a hybrid-AI ecosystem with broader applicability, in which descriptors will be designed to include (i) physical properties like atomic radius and valence electrons number, (ii) local geometric information, (iii) qualitative domain knowledge of interatomic interactions in terms of physical models, (iv) quantitative domain-knowledge of chemical bond-formation from DFT, and (v) quantitative domain knowledge on the structure-energy relation. These descriptors will be used with different regression models in combination with dimensionality reduction, hyper-optimization and importance analysis. This enables us to construct hybrid-AI models that are sufficiently robust to explore chemically and structurally complex topologically close-packed phases that are not accessible otherwise. The predictions of the hybrid-AI on structural stability and sublattice occupancy of these intermetallic materials will be confirmed by experimental measurements within this project. We expect that this hybrid-AI for materials will initiate new promising directions of research for the benefit of materials science in France and Germany.