Data-driven magnetic materials inverse design
Citation:
Cobelli, Matteo, Data-driven magnetic materials inverse design, Trinity College Dublin, School of Physics, Physics, 2024Download Item:
Abstract:
Magnetic materials have diverse applications across multiple sectors, ranging from magnetic resonance imaging machines, used to detect diseases, to electric motors, sensors, and wind turbines just to name a few. The demand for novel magnetic materials, tailored for specific applications, is higher than ever. However, the recent advances in technology have not been matched by a comparable rate of material discovery, largely due to the unavoidable low throughput of experimental synthesis. For these reasons, there is a growing need for alternative approaches to material discovery, potentially involving in-silico predictions. The modelling capabilities of Density Functional Theory (DFT) make it a promising technique for selecting material prototypes based on computed properties, leading to an inverse-design approach, where the material synthesis is driven by specific application needs. However, the computational cost of DFT is too high to be the only technique adopted for inverse-design purposes. In this work, we present a data-driven approach to the design of magnetic materials. We address various challenges that afflict the material-discovery pipelines, for which we find solutions that leverage recent advancements in artificial intelligence. The result is an end-to-end workflow for materials inverse-design with a strong interdisciplinary nature, borrowing techniques from various domains of machine learning, ranging from statistical modelling to natural language processing (NLP). Specifically, we present several novel methods. Firstly, we introduce an NLP pipeline for the automatic extraction of data from the scientific literature, based on the fine-tuning of transformers-based language models. We then employ a machine-learning-enhanced prototype generation technique to improve the creation of accurate convex hulls for assessing the stability of ternary alloys. Additionally, we have designed a local inversion algorithm for finding the atomic structure associated with a given set of atomic descriptors, that can be coupled with generative models. Finally, we introduce the Jacobi-Legendre potential, a linear machine-learning interatomic potential based on the cluster expansion of the system energy, as well as the spin power spectrum, a set of descriptors of the local chemical environment for the modelling of magnetic materials using machine-learning techniques.
Sponsor
Grant Number
Irish Research Council (IRC)
Author's Homepage:
https://tcdlocalportal.tcd.ie/pls/EnterApex/f?p=800:71:0::::P71_USERNAME:MCOBELLIDescription:
APPROVED
Author: Cobelli, Matteo
Advisor:
Sanvito, StefanoPublisher:
Trinity College Dublin. School of Physics. Discipline of PhysicsType of material:
ThesisAvailability:
Full text availableKeywords:
Magnetism, Machine Learning, Natural Language ProcessingMetadata
Show full item recordLicences: