Bandgap prediction for molecular crystals using Geometric Deep learning

Pathapati, Avinash, A (2020) Bandgap prediction for molecular crystals using Geometric Deep learning. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text s3754715_gnn_bandgap_final_report_upload_version.pdf Download (1MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (98kB)

Abstract

Deep Learning (DL) has achieved the state of the art results on many machine learning tasks starting from image classification, video processing to natural language understanding and speech recognition. The input data to the model in all of these tasks lie in the Euclidean space. However, there are many applications where the input data is not present in Euclidean space and are represented using graphs. For example, in chemistry, molecules are represented as graphs. Graph Neural Networks (GNNs), i.e., Geometric Deep Learning concerns generalized convolution methods that can work on a non-linear structure like graphs. In this thesis, we have examined the performance of GNNs and SchNet (a continuous filter convolution) model on OMDB dataset. The goal of the model is to predict the bandgap value given the structure of the molecule. Among all of the models tested, we found that our slightly modified version of SchNet achieved the Mean Absolute Error (MAE) of 0.28eV which is better than the state of the art. In this SchNet model, we also identified the atoms that are significant in contributing to the bandgap value of the molecule. Finally, we built an ensemble of SchNet models which attained a slightly lower MAE of 0.268eV. Although the GNN approaches that were tried did not improve the estimation accuracy, they still hold a promise in terms of improved explainability of results due to the graph-based nature of molecules.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Schomaker, L.R.B.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	28 Aug 2020 13:04
Last Modified:	28 Aug 2020 13:04
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/23263

Actions (login required)

View Item