# A Transfer Learning Approach for Microstructure Reconstruction and Structure-property Predictions

**Authors:** Xiaolin Li, Yichi Zhang, He Zhao, Craig Burkhart, L Catherine Brinson, Wei Chen  
**arXiv:** [1805.02784](https://arxiv.org/abs/1805.02784)  
**Date:** 2018-05-08  
**Summary page:** https://www.summarizepaper.com/en/paper/1805.02784v1/

## Key points

- Stochastic microstructure reconstruction plays a crucial role in computational materials science
- Current developments in this area are often limited to specific material systems
- The authors propose a transfer learning-based approach for microstructure reconstruction and structure-property predictions that can be applied to a wide range of material systems
- The proposed approach utilizes an encoder-decoder process and feature-matching optimization using a deep convolutional network
- Model pruning is implemented to study the correlation between microstructural features and hierarchical layers within the network
- The knowledge gained from model pruning is leveraged in developing a structure-property predictive model, determining the network architecture and initialization conditions
- Numerical experiments on various material microstructures with varying levels of complexity demonstrate the generality of the approach
- This approach offers an off-the-shelf solution for handling complex microstructures, unlike previous approaches that are limited or require extensive prior knowledge
- The approach has the potential to accelerate the discovery of new materials

## Summary

In the field of computational materials science, <kwd>stochastic microstructure reconstruction</kw>, plays a crucial role. However, current developments in this area are often limited to specific material systems. To address this issue, the authors propose a <kwrd>transfer learning-based approach</kw>, for microstructure reconstruction and structure-property predictions that can be applied to a wide range of material systems. The proposed approach utilizes an encoder-decoder process and feature-matching optimization using a deep convolutional network. For microstructure reconstruction, the authors implement <kwrd>model pruning</kw>, to study the correlation between microstructural features and hierarchical layers within the network. The knowledge gained from model pruning is then leveraged in developing a structure-property predictive model, which determines the network architecture and initialization conditions. To demonstrate the generality of their approach, the authors conduct numerical experiments on various material microstructures with varying levels of complexity in terms of geometrical characteristics. Unlike previous approaches that are limited to specific material systems or require extensive prior knowledge for model selection and hyper-parameter tuning, this approach offers an off-the-shelf solution for handling complex microstructures. Furthermore, it has the potential to accelerate the discovery of new materials. Overall, this paper presents a transfer learning-based approach that addresses the generality problem in stochastic microstructure reconstruction. By incorporating model pruning and leveraging obtained knowledge for structure-property predictions, this approach provides a versatile solution applicable to different material systems. Its ability to handle complex microstructures without extensive prior knowledge makes it a promising tool for accelerating materials discovery efforts.