Machine-learning accelerated materials discovery and optimisation
Prof. Patrick Rinke
Department of Physics, Technical University Munich, Germany
16:15 - 17:15 Tuesday 08 April 2025 TUG Materials are the foundation of technological advancements that shape our modern society. Their con-tinuous development enables new applications and products, while the discovery of novel materials addresses key societal challenges like clean energy production, sustainability, global prosperity, health, and wellbeing. Bottom-up materials development has traditionally relied on atomistic modelling using computational methods like density-functional theory (DFT). However, these methods are constrained by their high computational costs. In this talk, I will present different ways in which machine-learning can overcome such computational bottlenecks and provide new insight and significant acceleration.
For the new family of mixed metal chalcohalides that are considered for photovoltaic applications, I will illustrate how data science methods accelerate novelty validation [1]. I will then show how ex-plainable machine-learning methods help us to derive design rules from DFT data that facilitates further materials optimization.
Next, I will introduce machine-learned interatomic potentials (MLIPs), such as MACE [2], that are cur-rently revolutionizing atomistic modelling. MLIPs learn the interaction between atoms from, e.g., DFT calculations and can then simulate atomistic systems at much reduced cost and therefore increased complexity. We have developed active learning workflows to train MACE MLIPs efficiently and accu-rately for different applications. I will illustrate for halide and hybrid perovskites considered for pho-tovoltaic applications, that MLIPs facilitate the investigation of perovskite alloys and allow us to opti-mize materials parameters like the band gap [3]. Even quaternary alloys are now within reach, and we can assess the viability of stabilizing lead-free, tin-based perovskites with organic components. In the context of catalysis, we have used MLIPs from the Open Catalyst Project (OCP [4]) to compute nearly a million adsorption energies of catalytically relevant molecules on different metals surfaces. We use the corresponding adsorption energy distributions as descriptors for catalytically interesting systems and propose new catalysts for CO2 to methanol conversion [5]. Lastly, I will demonstrate how we ex-tend MLIP training to infrared (IR) spectra and demonstrate that MLIPs can provide accurate IR spectra for catalytically relevant molecules at a fraction of the computational cost of the conventional DFT IR approach.
1. Screening Mixed-Metal Sn2M(III)Ch2X3 Chalcohalides for Photovoltaic Applications, P. Henkel, J. Li, G. Krish-namurthy Grandhi, P. Vivo, and P. Rinke, Chem. Mat. 35, 18, 7761 (2023)
2. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields, I. Batatia, D. P. Kovacs, G. Simm, C. Ortner, and G. Csányi, Adv. Neural. Inf. Process. Syst. 35, 11423 (2022)
3. Compositional engineering of perovskites with machine learning, J. Laakso, M. Todorovic, J. Li, G.-X. Zhang and P. Rinke, Phys. Rev. Materials 6, 113801 (2022)
4. Open Catalyst Project: https://opencatalystproject.org/
5. Machine-learning Accelerated Descriptor Design for Catalyst Discovery: A CO2 to Methanol Conversion Case Study, P. Pisal, O. Krejci, and P. Rinke, arXiv:2412.13838v2
|
