More pictures here!
Symposium SB11: Bio-based and Biomimetic Polymers in Soft Robotics

Symposium MT03: Machine Learning Methods, Data and Automation for Sustainable Electronics

Victor Fung, Georgia Institute of Technology

Physics-Informed Pre-training of Graph Neural Networks for Materials Property Predictions

Written by Kazi Zihan Hossain

Materials scientists have used machine learning models and artificial intelligence (AI) to predict the properties of different materials. Traditional approaches require individual models and re-training for every different system to predict different properties. A model trained to predict a specific property can not be used to predict another property of the materials. To overcome this issue and develop a single model that can be used to predict different properties of materials, Victor Fung from the Georgia Institute of Technology has presented a Graph Neural Network (GNN) approach. In this technique, the composition and structure of a material can be encoded into a graph so that the properties of the materials can be related to the structure. Traditionally, GNN approaches work well when enough data is available to train the model, which may not be feasible from a material science perspective. Therefore, Fung and colleagues used the transfer learning technique to pre-train the model from a large dataset and then used the model to tune the properties of the materials. Different models, such as Crystal Graph Convolutional Neural Networks (CGCNN) and TorchMD-Net, were used to predict the properties of the materials, where the latter performed well under denoising scenarios.

Comments

The comments to this entry are closed.