NestedAE: interpretable nested autoencoders for multi-scale materials characterization.

Area of Science:

• Materials Science
• Machine Learning
• Data Science

Background:

• Multiscale materials present complex challenges for data analysis due to varying property scales.
• Interpretable machine learning models are crucial for understanding structure-property-performance relationships.

Purpose of the Study:

• Introduce NestedAE, a novel interpretable machine learning architecture for analyzing multiscale materials.
• Benchmark NestedAE against traditional autoencoders (AE) for performance and robustness.
• Investigate the relationship between crystal-scale properties and device performance using real-world data.

Main Methods:

• Developed NestedAE, a supervised autoencoder architecture with nested structures.
• Validated NestedAE on a synthetic dataset with known dimensionality.
• Applied NestedAE to a multiscale MHP dataset combining atomic/ionic properties and device J-V characteristics.

Main Results:

• NestedAE demonstrated superior noise robustness and lower reconstruction losses than vanilla AE.
• The model successfully identified links between crystal structure properties and device performance.
• Results align with existing experimental observations in multiscale materials.

Conclusions:

• NestedAE offers a powerful and interpretable approach for multiscale materials analysis.
• The architecture effectively bridges the gap between fundamental material properties and macroscopic device behavior.
• This work facilitates deeper understanding and prediction in advanced materials development.