Multitask deep learning with dynamic task balancing for quantum mechanical properties prediction.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Predicting quantum mechanical properties (QMPs) is crucial for advancing material and chemistry science.
• Multitask deep learning models are prevalent for QMPs prediction but often overlook task interdependencies, leading to overfitting.
• Existing models lack mechanisms to dynamically adjust for varying task difficulties.

Purpose of the Study:

• To develop an improved multitask learning framework for accurate QMPs prediction.
• To address the overfitting issue in multitask learning by considering task relationships and difficulties.
• To enhance the interpretability of deep learning models in predicting molecular orbital energy levels.

Main Methods:

• Proposed a multiscale dynamic attention graph neural network (MDGNN) for molecular representation learning.
• Introduced a dynamic task balancing (DTB) strategy to manage task differences and difficulties.
• Utilized gradient-weighted class activation mapping (Grad-CAM) for model interpretability, specifically for frontier molecular orbital predictions.

Main Results:

• The MDGNN demonstrated superior performance compared to state-of-the-art multitask learning models on two large QMPs datasets.
• The DTB strategy significantly improved the performance of the MDGNN.
• Grad-CAM analysis provided explanations consistent with molecular orbital theory, validating the model's interpretability.

Conclusions:

• The MDGNN with DTB strategy offers improved generalization and interpretation capabilities for QMPs prediction.
• This approach advances the application of deep learning in computational chemistry and materials science.
• The findings pave the way for more reliable and understandable predictive models in the field.