Modeling the interpretable geometric-performance relationship of metamaterials on small datasets using...

Area of Science:

• Materials Science
• Computational Physics
• Artificial Intelligence

Background:

• Deep learning methods for metamaterial property prediction often lack interpretability and require large datasets.
• Existing approaches are computationally expensive and hinder the exploration of underlying physical mechanisms.
• Multi-layer perceptron-kernelled methods present opacity in end-to-end mapping.

Purpose of the Study:

• To introduce a novel, lightweight neural network architecture for metamaterial property prediction.
• To develop a framework for dataset construction, high-fidelity analysis, and performance visualization.
• To enable accurate prediction and exploration of physical mechanisms using small datasets.

Main Methods:

• Proposed the Kolmogorov-Arnold Operator Informed Network (KAOIN), a novel neural network structure.
• Developed a coupled metamaterial performance prediction framework.
• Validated KAOIN using spatial symmetry and the Gibson-Ashby theoretical model for interpretability.

Main Results:

• KAOIN achieved improved accuracy and convergence speed under small-sample conditions.
• The framework successfully predicted specific energy absorption for gradient triply-periodic minimal surfaces using only 50 data points.
• Incorporating a geometry-performance relationship enhanced accuracy by up to 44.6% and convergence speed by 48-89%.

Conclusions:

• KAOIN offers an interpretable neural network paradigm for exploring physical mechanisms with limited data.
• The developed framework facilitates accurate modeling of metamaterial geometric-performance relationships.
• This approach shows significant potential for advancing metamaterial design and discovery.