Microstructure Representation and Reconstruction of Heterogeneous Materials via Deep Belief Network for Computational Material Design (1612.07401v3)

Abstract: Integrated Computational Materials Engineering (ICME) aims to accelerate optimal design of complex material systems by integrating material science and design automation. For tractable ICME, it is required that (1) a structural feature space be identified to allow reconstruction of new designs, and (2) the reconstruction process be property-preserving. The majority of existing structural presentation schemes rely on the designer's understanding of specific material systems to identify geometric and statistical features, which could be biased and insufficient for reconstructing physically meaningful microstructures of complex material systems. In this paper, we develop a feature learning mechanism based on convolutional deep belief network to automate a two-way conversion between microstructures and their lower-dimensional feature representations, and to achieves a 1000-fold dimension reduction from the microstructure space. The proposed model is applied to a wide spectrum of heterogeneous material systems with distinct microstructural features including Ti-6Al-4V alloy, Pb63-Sn37 alloy, Fontainebleau sandstone, and Spherical colloids, to produce material reconstructions that are close to the original samples with respect to 2-point correlation functions and mean critical fracture strength. This capability is not achieved by existing synthesis methods that rely on the Markovian assumption of material microstructures.

• The paper introduces a Convolutional Deep Belief Network (CDBN) for automating the representation and reconstruction of heterogeneous material microstructures within a low-dimensional space.
• The proposed CDBN method achieves a 1000-fold dimensionality reduction and successfully reconstructs microstructures of diverse materials, maintaining key features and properties like critical fracture strength.
• This deep learning approach offers a powerful tool for overcoming challenges in high-dimensional microstructure design, potentially simplifying structural design and enhancing processing-structure-property mappings in computational materials engineering.

Microstructure Representation and Reconstruction of Heterogeneous Materials via Deep Belief Network for Computational Material Design

The paper presents an innovative approach to microstructural representation and reconstruction of heterogeneous materials using Convolutional Deep Belief Network (CDBN) within the framework of Integrated Computational Materials Engineering (ICME). The objective is to address existing constraints in microstructure design and reconstruction processes, specifically within high-dimensional image spaces.

The researchers offer a compelling case for utilizing deep learning techniques to automate feature extraction from complex material systems. By adopting CDBN, a two-way conversion between microstructures and their lower-dimensional feature representations is streamlined, achieving a notable dimensionality reduction—specifically, a 1000-fold reduction in the microstructure space. This breakthrough is posited as an advancement over conventional approaches that require human interpretation and manual extraction of geometric and statistical features, which can be subjective and limited.

A significant portion of the paper is dedicated to detailing the architecture and parameters of the proposed CDBN model. The model configurations adopted, such as the filters, sparsity constraints, and probabilistic max-pooling layers, are thoroughly optimized to replicate complex microstructure features across multiple material systems. These systems include Ti-6Al-4V alloy, Pb63-Sn37 alloy, Fontainebleau sandstone, and spherical colloids. The goal is to produce reconstructions that closely align with original samples, assessed through 2-point correlation functions and mean critical fracture strengths.

The results indicate successful reconstruction across diverse material types with significant visual fidelity and maintenance of crucial material properties. Moreover, the statistical match in critical fracture strength between original and reconstructed samples suggests that such features could potentially simplify structural design, enhancing predictability in processing-structure-property mappings within ICME.

Despite its achievements, the research acknowledges limitations such as the non-scalability of the reconstruction process, where microstructure samples are bound to fixed dimensions. Additionally, discrepancies between original samples and their reconstructions concerning fine-grained details are noted, alongside the computational difficulty in feature extraction network tuning. These issues represent critical avenues for future exploration, particularly the integration of stochastic models to enable scalable synthesis.

In conclusion, the paper provides valuable insights into the role of automated deep learning models in revolutionizing material design, presenting a powerful tool for addressing challenges of high-dimensional microstructure representation. The methodology shared could significantly influence future theoretical and practical applications within computational material design, particularly with the potential incorporation of scalable synthesis models and enhanced network training strategies. The investigation into maintaining property fidelity post-reconstruction also offers a crucial facet for ongoing paper and potential validation of physical feasibility.