%0 Thesis
%A Hilgers, Robin
%T Prediction of Magnetic Materials for Energy and Information Combining Data-Analytics and First-Principles Theory
%V 288
%I RWTH Aachen University
%V Dissertation
%C Jülich
%M FZJ-2024-06074
%@ 978-3-95806-795-0
%B Reihe Schlüsseltechnologien / Key Technologies
%P xv, 215
%D 2024
%Z First published with RWTH Aachen University
%Z Dissertation, RWTH Aachen University, 2024
%X The essential role of magnetic materials in information technology and the corresponding energy consumption of data storage centers is crucially underestimated in modern society. Saving energy resources is the societal challenge of the 21st century. One of the leading scientific objectives is finding ways to reduce energy consumption and make resource usage more efficient. This thesis aims to shed light on possible contributions of materials science simulations towards a green IT transformation by providing workflows and best-practice guidelines for high-throughput materials screening tasks. An instance of such a screening task is the search for magnetic materials for the next generation of storage and data processing devices. However, as the simulation process itself is time-consuming, this thesis explores not only the material phase space but also the application opportunities for data science and machine learning (ML) in the material’s property prediction process. As a prime example of a complex magnetic material property, which is a limiting quantity when it comes to methodological applicability, the critical temperature 𝑇𝑐 of existing magnetic simulation data of Heusler alloys will be predicted using ML models. The capability and limitations of these models will be analyzed and discussed. It is shown that it is possible to extract physical relations and knowledge from trained ML models without any prior knowledge of the underlying physics and system mechanics. Whether a Heusler compound has a 𝑇𝑐 high enough to be relevant for an application in magnetic data storage and processing devices could be predicted with over 90 % accuracy using lightweight ML model algorithms on typical materials science data set sizes. Beyond that, the phenomenon of near half-metallicity in Heusler compounds was examined, including the successful ML-based prediction of compounds displaying this property which were not known to be nearly half-metallic before (L21 Co2HfIn, XA Mn2TaGe, and L21 Co2ScSn). This particular study used existing first-principles data of full and inverse Heusler compound’s spin-polarized density of states, in order to screen publicly available structural and magnetic ab initio data for compounds exhibiting near half-metallic properties. The relations learned by the underlying ML models are discussed and compared to a known physical model. It was determined that ML models have the capability to extend and complement known physical models and relations when applied to existing (and potentially imperfect) data. Finally, large-scale high-throughput ultrathin film simulations of 3𝑑 transition metal layers on face-centered cubic noble metal substrates were performed to understand the magnetic properties of these magnetic multilayer films, which are predicted to represent well-suited host platforms for room temperature stable Skyrmions and hence are considered candidate materials for spintronics-based storage and data processing device applications. Tailored to high-throughput ab initio workflows, a scalable method—that increased the overall convergence rate from 64.8 % to 94.3 % and exhibited the potential to save up to 17 % of the computational time required, as well as to reduce the number of needed ab initio relaxation steps to relax a multilayer film system by up to 29 % in this systematic study, while being flexible enough also to be applicable to future use cases—using the integration of batch learning into high-throughput workflows, was developed. The use, restrictions, implementation, starting conditions, and benefits of ML-based techniques and explainable artificial intelligence are discussed in depth in this thesis.
%F PUB:(DE-HGF)3 ; PUB:(DE-HGF)11
%9 BookDissertation / PhD Thesis
%R 10.34734/FZJ-2024-06074
%U https://juser.fz-juelich.de/record/1032226