September 19, 2024
11am – 12.15pm : Rotation-equivariant graph neural networks for learning glassy liquids representations
Speaker: Francois P. Landes, Laboratoire Interdisciplinaire des Sciences du Numérique (LISN), Université Paris-Saclay
Abstract: In this talk, I will motivate the development and use of group-equivariant Neural-Networks (a.k.a. "geometric deep learning"), which cover a broad range of applications. I will focus on the case of SE(3)-equivariant GNNs (group of roto-translation) applied to glassy systems.
After quickly introducing the task (learning a representation of a glassy liquid) and sketching out the key features of Graph Neural Networks (esp. node-equivariance), I will insist on the notions of rotation equivariance and invariance, and how to embed them in a network [Thomas 2018].
I will then present our specific architecture [Pezzicoli 2023], outline our most salient results (reduced number of parameters, improved generalization and interpretability) and open with perspectives on more recent architectures.
[Thomas 2018: Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds]
[Pezzicoli 2023: Rotation-equivariant graph neural networks for learning glassy liquids representations]
12.15 pm - 1.45 pm : Lunch break
1.45 pm – 3 pm: Combining the Power of High-Throughput Ab Initio Calculations and Machine Learning towards Materials Informatics
Speaker: Gian-Marco Rignanese, Université catholique de Louvain, Belgique (UCLouvain)
Abstract: The progress in first-principles simulation codes and supercomputing capabilities have given birth to the so-called high-throughput (HT) ab initio approach, thus allowing for the identification of many new compounds for a variety of applications (e.g., lithium battery and photovoltaic). As a result, a number of databases have also become available online, providing access to various properties of materials, mainly ground‑state though. Indeed, for more complex properties (e.g., linear or higher‑order responses), the HT approach is still out of reach because of the required CPU time. To overcome this limitation, machine learning approaches have recently attracted much attention in the framework of materials design.
In this talk, I will review recent progress in the emerging field of materials informatics. I will briefly introduce the OPTIMADE API [1] that was developed for searching the leading materials databases (such as AFLOW, the Materials Cloud, the Materials Project, NOMAD, OQMD, ...) using the same queries. I will introduce the MODNet framework and its recent developments for predicting materials properties [2-5]. This approach, which is particularly well suited for limited datasets, relies on a feedforward neural network and the selection of physically meaningful features. Finally, I will illustrate the power of materials informatics, combining high‑throughput ab initio calculations and machine learning, through a few recent examples.
References
[1] C.W. Andersen et al., Sci. Data 8, 217 (2021).
[2] P.‑P. De Breuck, G. Hautier, and G.‑M. Rignanese, npj Comput. Mater. 7, 83 (2021).
[3] P.‑P. De Breuck, M.L. Evans, and G.‑M. Rignanese, J. Phys.: Condens. Matter 33, 404002 (2021).
[4] P.‑P. De Breuck, G. Heymans, and G.‑M. Rignanese, J. Mater. Inf. 2, 10 (2022).
[5] X. Liu, P.‑P. De Breuck, L. Wang, and G.‑M. Rignanese, npj Comput. Mater. 8, 233 (2022).
