Special Seminar

Combined DFT and ML for Materials Design

Mohammadreza Karamad, SFU Chemistry
Location: P8445.2

Friday, 01 March 2024 11:30AM PST
Facebook
Twitter
LinkedIn
Reddit
SMS
Email
Copy

Synopsis

The global energy landscape must undergo significant transformation to sustainably meet our long- term energy demands. Addressing the challenge of reducing CO2 emissions from fossil fuels requires access to safe, clean, and sustainable energy sources. An important aspect of advancing clean energy technologies lies in the facile discovery of novel materials tailored to these applications. Hydrogen is emerging as a globally recognized alternative to fossil fuels for clean energy technologies, but challenges remain in its low-cost, environmentally friendly distribution. Hydrogen storage in solid state materials is a promising approach. However, current solid-state materials such as chemical hydrides face limitations such as on-board regeneration costs. Advancements in ab initio calculations such as density functional theory (DFT) as well as increasing computing power have enabled high-throughput computational screening to identify new materials with specific properties of interest a priori, presenting a promising avenue for addressing hydrogen storage challenges.

In addition to quantum mechanical calculations, Machine Learning (ML) techniques have proven to provide a fast and accurate way to predict desired properties enabling facile discovery of new materials. One of the key challenges to develop a ML-based approach for predicting materials properties is materials representation. Designing descriptors that can represent the local chemical environments of atoms is necessary to develop a robust ML model with high predictive accuracy.

In the first part of my presentation, I will discuss my efforts in designing hydrogen storage materials using DFT calculations. In the second part of my talk, I will discuss the recently developed graph convolutional neural network framework, Orbital Graph Convolutional Neural Network (OGCNN), for predicting properties of crystalline materials. This framework takes advantageous of powerful Convolutional Neural Networks (CNNs) enabling the extraction of spatial information from neighbouring atoms in the crystalline materials in conjunction with graph representation of materials with descriptors such as orbital-orbital interactions between constituent atoms.