Richard Hennig Jason Gibson - AI-driven Workflows for the Discovery of Novel Superconductors - 2023

Title : Richard Hennig Jason Gibson - AI-driven Workflows for the Discovery of Novel Superconductors Author(s): Institute for Pure & Applied Mathematics (IPAM) Link(s) : https://www.youtube.com/watch?v=UTbAVPveSMI

Relevant to AI for Science.

Computational material discovery in the last century involved:

• Experimental trial and error based on intuition and empirical rules.
• Exploration of material spaces near known compounds and alloys.

The new era:

• Computationally driven discovery of \(\text{Celr}_4\text{In}\) compound.
• Ferromegnatic 2D- \(\text{Fe}_3\text{GeTe}_2}\).
• Prediction of \(\text{WB}_2\) superconductor.

AI-Driven workflows:

1. Mining databses and literature.
2. Take candidates from above, predicting the crystal structure. Currently done via genetic algorithm GASP.
3. Predict properties - electron-phonon coupling, critical temperature and fields.
4. Get material stability - requires free energy calculations, phase diagrams.

GASP Algorithm:

• Generate structures (mutation and crossover).
• Local optimization through relaxation.
• Identify low-energy basins of attraction.

Speakers are adding a machine learning component to make specific predictions within the genetic algorithm loop. The ML model used is the Crystal Graph Neural Network (GNN), which predicts the energy, forces and formation energy. However, the results are worse on unrelaxed structures (#DOUBT Whats an unrelaxed structure?) compared to relaxed structures (order of magnitude higher errors). To solve this problems, they converted pertubed structures to what their relaxed structures would be which is a combination of step functions.

Material prediction and design requires energy landscape, which these days make use of Density Functional Theory, empirical potential molecular dynamics.

Flexible spline-based potentials are accurate, fast to fit and evaluate, and easy to interpret, see (, ).