Creating a Microstructure Latent Space with Rich Material Information for Multiphase Alloy Design

TL;DR

This work addresses the challenge of linking microstructure to composition, processing, and properties in multiphase alloys by introducing a microstructure-centered CPSP framework. A variational autoencoder–driven deep learning model (VAE–DLM) encodes microstructural images into a latent space and jointly predicts composition/processing and material properties, enabling CPSP inversion through latent-space sampling. The approach is demonstrated by designing UniDP steels and validating them experimentally, with the latent space shown to be continuous and information-rich, facilitating interpolation across microstructure, composition, and performance. The method offers a pathway for rapid, microstructure-informed alloy design with potential to streamline development and reduce experimental costs in multiphase systems.

Abstract

The intricate microstructure serves as the cornerstone for the composition/processing-structure-property (CPSP) connection in multiphase alloys. Traditional alloy design methods often overlook microstructural details, which diminishes the reliability and effectiveness of the outcomes. This study introduces an improved alloy design algorithm that integrates authentic microstructural information to establish precise CPSP relationships. The approach utilizes a deep-learning framework based on a variational autoencoder to map real microstructural data to a latent space, enabling the prediction of composition, processing steps, and material properties from the latent space vector. By integrating this deep learning model with a specific sampling strategy in the latent space, a novel, microstructure-centered algorithm for multiphase alloy design is developed. This algorithm is demonstrated through the design of a unified dual-phase steel, and the results are assessed at three performance levels. Moreover, an exploration into the latent vector space of the model highlights its seamless interpolation ability and its rich material information content. Notably, the current configuration of the latent space is particularly advantageous for alloy design, offering an exhaustive representation of microstructure, composition, processing, and property variations essential for multiphase alloys.

Figures (12)

• Figure 1: Microstructure-centered alloy design framework. (a) Preprocessing of literature-sourced microstructural images through binarization and data expansion. (b) Construction of a deep learning model comprising an encoder, a decoder, and dual multilayer perceptrons to establish the microstructure-centered composition/processing–structure–property (CPSP) connections. (c) Design and validation process involving random sampling within the latent vector space, guided by physical metallurgy principles, to identify potential unified dual-phase (UniDP) steel candidates.
• Figure 1: Details of data on dual-phase (DP) steels gathered from literature. (a), (b) Representative scanning electron microscope (SEM) images. (c) Diverse heat treatment process routes. Different process routes correspond to different coding serial numbers to facilitate model training.
• Figure 2: Architecture and predictive capabilities of the VAE-centric deep learning model (VAE–DLM). (a) Model structure including an encoder, a decoder, and two multilayer perceptrons (called the MCP and MP models). The encoder, adapted from ResNet18, processes binary images (ferrite in black, martensite in white) to produce a probability distribution characterized by the mean ($\mu$) and standard deviation ($\sigma$). The decoder reconstructs a similar image from a latent vector drawn from this distribution. Utilizing the latent vectors, the MCP model predicts the composition and process of the alloy while the MP model predicts its properties. (b) Predictive performance across all the outputs. The term “Ta” and “ta” denote the annealing temperature and annealing time. Comparison of predicted and actual values of (c) ultimate tensile strength (UTS) and (d) uniform elongation (UE), respectively.
• Figure 2: Mean absolute error for two series of machine learning models on ultimate tensile strength (UTS) and uniform elongation (UE). (a) The first series (named CPP method) that uses composition and process as inputs. (b) The second series (named CPMP method) that uses composition, process, and martensite volume fraction as inputs. The machine learning models include support vector regression (SVR) [1], random forest (RF) [2], multilayer perceptron (MLP) [3], gradient boosting regression (GBR) [4] and extreme gradient boosting (XGB) [5]. Notably, MLP exhibits the lowest mean absolute error in both series, respectively.
• Figure 3: UniDP steel design. (a) Mechanical property of the sampling points at $\varepsilon$$\sim$$\mathcal{N}$(0, $10^{2}$). (b) Compositional distribution of the alloy in the blue box in (a). (c) Mechanical properties of the experimental alloys. Designed new alloys meet the performance requirements of DP780, DP980, and DP1180, respectively. (d) Elemental cost comparison between the experimental and commercial alloys.