Message-passing neural networks for high-throughput polymer screening | JoVE Visualize

Area of Science:

Computational chemistry
Materials science
Machine learning

Background:

Machine learning (ML) models excel at predicting molecular properties, enabling rapid virtual screening.
Graph-based neural networks (GNNs) are state-of-the-art for molecular structure predictions.
Current GNNs often require computationally expensive 3D molecular geometries, limiting their use in high-throughput screening.

Purpose of the Study:

To develop and evaluate ML models for predicting properties of large molecules relevant to organic photovoltaics.
To assess the necessity of 3D structural information for accurate predictions in large molecular systems.
To investigate the transferability of learned molecular representations for new computational chemistry tasks.

Main Methods:

Creation of a novel database of ~91,000 candidate molecules for organic photovoltaics, including large molecules (up to 200 atoms) and extrapolated polymer properties.
Training and evaluation of message-passing neural networks (MPNNs) with and without 3D structural inputs.
Analysis of prediction accuracy and comparison with existing benchmark datasets and state-of-the-art methods.
Exploration of learned molecular representations for transfer learning across different density functional theory (DFT) functionals.

Main Results:

MPNNs trained with and without 3D information achieved comparable accuracy for the large molecules in the new dataset.
Prediction accuracy was on par with state-of-the-art methods on existing benchmark datasets.
Learned molecular representations facilitated reduced training data requirements for transferring predictions to new DFT functionals.

Conclusions:

Optimized 3D molecular geometry is not essential for achieving high accuracy in ML predictions for large molecules relevant to organic photovoltaics.
The developed dataset and ML approach offer a practical and efficient method for high-throughput screening of novel organic photovoltaic materials.
Learned molecular representations show promise for efficient model adaptation and transfer learning in computational chemistry.