報 告 人：Miguel A.L. Marques

報告題目：Accelerating the discovery and characterization of new inorganic materials with machine learning

報告時間：2024年10月16日（周三）上午10：00

報告地點：物理與電子工程學院428會議室

主辦單位：物理與電子工程學院、科學技術研究院

報告人簡介：

Prof. Miguel A.L. Marques received his PhD degree in Physics from the University of Würzburg in 2000, under the supervision of E.K.U. Gross, with a thesis in the field of density functional theory for superconductors. He then held several post-doctoral positions in Spain, Germany, and in France. From 2005 to 2007 he was assistant professor at the University of Coimbra in Portugal. From 2007 to 2014 he was a CNRS researcher (CR1) at the University of Lyon 1. Since then he is a professor at the Martin-Luther University of Halle-Wittenberg, and now is the leader of the research group of Artificial intelligence for intergrated material science in Ruhr University Bochum . His current research interests include density functional theory, superconductivity, application of machine learning to materials science, materials for energy applications, etc. He was the initial developer of octopus, a code that is currently used by hundreds of researchers worldwide to study ab-initio dynamics of electrons, and is the main developer of libxc, a library of exchange-correlation functionals that is used by more than 30 codes in quantum chemistry and condensed matter physics. He authored more than 310 articles (with a total of more than 24000 citations) and has a Hirsch index of 56 (source: Google Scholar). He also edited three books published by Springer in their Lecture Notes in Physics series, edited two special issues in International Journals, and translated three books of popular science to Portuguese. He organized several summer schools and international workshops, the most relevant of which are the series of the Benasque School and International Workshop in TDDFT, that takes place in Benasque, Spain every second year, and was chair of the 2009 Gordon conference on TDDFT. He is the spokesperson of the PsiK network for the fields of GW and TDDFT and member of its advisory scientific committee, and Research Team leader in the European Theoretical Spectrocopy Facility (ETSF).

報告摘要：

We summarize our recent attempts to discover, characterize, and understand inorganic compounds using novel machine learning approaches. We start by motivating why the search for new materials is nowadays one of the most pressing technological problems. Then we summarize our recent work in using crystal-graph attention neural networks for the prediction of materials properties. To train these networks, we developed a dataset of over 4.5 million density-functional calculations with consistent calculation parameters. Combining the data and the newly developed networks we have already scanned thousands of structural prototypes spanning a space of several billion materials and identified tens of thousands of theoretically stable compounds. We then discuss how simple, interpretable machine learning approaches can be used to understand complex material properties, such as the transition temperature of superconductors. Finally, we speculate which role machine learning will have in the future of materials science.