Simultaneously improving accuracy and computational cost under parametric constraints in materials property...

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning in Materials

Background:

• Machine learning (ML) and deep learning (DL) are effective for materials property prediction.
• Current deep neural network (DNN) models often require extensive computational resources and long training times.
• There is a need for efficient DNNs that balance predictive accuracy with computational feasibility for materials research.

Purpose of the Study:

• To propose a novel deep neural network framework, iBRNet, for regression-based materials property prediction.
• To develop a DNN architecture that reduces model parameters and training time without sacrificing accuracy.
• To evaluate iBRNet's performance against traditional ML and existing DL models across various datasets.

Main Methods:

• Developed iBRNet, a deep regression neural network featuring branched skip connections and multiple schedulers.
• Utilized composition-based numerical vectors representing elemental fractions as input for model training.
• Compared iBRNet's performance against state-of-the-art ML and DL architectures on multiple datasets.

Main Results:

• iBRNet consistently outperformed traditional ML and known DL models in materials property prediction across all dataset sizes.
• The branched structure and multiple schedulers in iBRNet led to a reduction in model parameters.
• iBRNet demonstrated faster model training times and improved convergence compared to other neural network architectures.

Conclusions:

• The proposed iBRNet framework offers an efficient and accurate solution for regression-based materials property prediction.
• Combining multiple callback functions within DNNs effectively minimizes training time and maximizes accuracy under computational constraints.
• iBRNet provides a valuable tool for researchers, enabling advanced materials property prediction with reduced computational demands.