Rational Design of Deep Learning Networks Based on a Fusion Strategy for Improved Material Property Predictions.

Area of Science:

• Materials Science
• Machine Learning
• Computational Chemistry

Background:

• Deep learning model design in materials science often lacks rational guidance, relying heavily on trial and error.
• Existing models have limitations, necessitating fusion strategies for improved performance and expanded design possibilities.

Purpose of the Study:

• To address the limitations of current deep learning models in materials science by introducing a physically insightful feedback method.
• To develop a novel deep learning model for structure-agnostic material property prediction that leverages enhanced design strategies.

Main Methods:

• Developed the Chemical Environment Clustering Vector (CECV) as a physically insightful feedback mechanism.
• Designed a Long Short-Term Memory and Gated Recurrent Unit fused with Deep Convolutional Neural Network (L-G-DCNN) model based on CECV.
• Applied the L-G-DCNN model to structure-agnostic material property predictions.

Main Results:

• The L-G-DCNN model accurately captures element interactions within compounds, leading to precise material property predictions.
• L-G-DCNN outperformed state-of-the-art structure-agnostic models across 28 benchmark datasets.
• The model demonstrated superior sample efficiency and faster convergence speeds compared to existing methods.

Conclusions:

• The CECV-based fusion strategy significantly improves the comprehension and design of deep learning models in materials informatics.
• This approach offers a novel perspective for advancing materials informatics research through enhanced model design and prediction accuracy.