The Office of Dean-Research is organising the 23rd edition of the University Distinguished Lecture inviting Prof. Stefano Sanvito, Professor and Chair of Condensed Matter Theory, School of Physics and Director, CRANN (Centre for Research on Adaptive Nanostructures and Nanodevices), Trinity College Dublin, Ireland.
Title of the Lecture: “Is AI Really Useful for Materials Science?”
Abstract
AI and machine learning have rapidly become important tools for our daily life, providing assistance over many tasks. These, however, often concern relatively low-stake domains, such as idea generation, text summarisation, automated paper writing/ reviewing, etc. The physical sciences, however, demand far more: accuracy at the level of experimental validation, reproducibility across thousands of steps and the orchestration of a vast and diverse toolset.
In this talk I will describe our attempts at integrating machine-learning and AI methods into materials science, by looking at rather different domains. Firstly, I will discuss a machine-learning scheme for predicting the Curie temperature of ferromagnets, which uses solely the chemical composition of a compound as feature and experimental data as target [1]. In particular, I will discuss how to develop meaningful feature attributes for magnetism and how these can be informed by experimental and theoretical results. Furthermore, I will show how the experimental data can be mined from published scientific literature with the help of natural language processing tools [2].
Then, I will describe how an accurate description of the structure of materials, which is amenable to be used with machine learning, can offer a quantum-chemistry-accurate description of local properties at virtually no computational costs. This is based on a newly developed cluster expansion formulated in terms of Jacobi-Legendre polynomials. Such general framework allows us to construct, on the same footing, extremely accurate force fields [3] as well as converged electron densities [4]. These can both be integrated in material-discovery workflows generating unprecedented speedup.
Finally, I will present two examples of how visual-graphic techniques can be integrated into materials science. In the first case autoencoder neural networks will be used to denoise scanning electron transmission microscopy images [5], thus allowing us to perform microscopy on fragile samples, which rapidly degrade under electron irradiation. In the second case, I will use movie- enhancement methods to provide high-resolution tomographic maps of complex materials complexes [6]. Importantly, the same method can also be used for medical imaging and can be incorporated in medical devices.
[1] Physical Review Materials 3, 104405 (2019).
[2] npj Computational Materials 9, 222 (2023).
[3] Physical Review B 108, 094102 (2023).
[4] npj Computational Materials 9, 87 (2023).
[5] Machine Learning: Science and Technology 4, 015025 (2023).
[6] Nature Communications 15, 7962 (2024).