In the past two months, Professor Xu Wei’s team at the State Key Laboratory of Digital Steel, NEU, published four high-quality research papers across three consecutive issues of Acta Materialia, focusing on systematic methodologies in the field of “AI for Microstructure”. The principal authors include Professor Xu Wei, Associate Professor Wang Chenchong, and Distinguished Associate Research Fellow Wei Xiaoliao, with doctoral students Ma Xudong and Han Siyu also contributing significantly to the publications.
In recent years, “AI for Science” has been recognized as the fifth scientific paradigm—following the empirical, theoretical, computational, and data-driven models—ushering in a new era of paradigm shifts and deeper interdisciplinary integration. In the field of materials science, artificial intelligence offers unprecedented opportunities to accelerate the prediction of material properties and the design of novel materials—showing exceptional promise, in particular, for the development of complex metallic systems such as advanced steels. However, current AI for Science approaches still primarily focus on modeling structured data related to composition, processing, and properties. In the “composition/process–microstructure–property” relationship, microstructure—the most critical intermediate link—remains largely underexplored by AI, with a notable lack of dedicated methodologies.
To tackle this issue, Professor Xu Wei’s team at NEU has systematically developed a series of AI methodologies under the theme “AI for Microstructure”. These methodologies are grounded in a variety of advanced AI architectures, including large-scale semantic segmentation models for microstructure analysis, self-supervised generative frameworks, multimodal convolutional neural networks, and physics-informed transfer learning models. Collectively, they support a broad spectrum of tasks, such as generalized microstructure recognition for metallic structural materials, microstructure image prediction, inverse analysis of mechanical behaviors, complex property prediction, and alloy design. These efforts encompass the following components:
1. Microstructure Recognition with Large Image Models
Accurate recognition and quantification of microstructures are critical to understanding the properties of metallic materials. Although supervised learning and fine-tuning of large-scale models have attracted widespread attention, they both depend heavily on labeled datasets and additional training tailored to specific tasks. The Segment Anything Model (SAM), a prominent foundation model in the field of natural image processing, offers strong zero-shot segmentation capabilities, enabling rapid instance segmentation. Professor Xu Wei’s team has pioneered a general algorithm for phase segmentation of alloy microstructures by integrating the Segment Anything Model (SAM) with domain knowledge-guided post-processing—a first in the field. The algorithm leverages SAM to produce initial segmentation outputs, which are subsequently refined through standardized rules informed by metallurgical domain expertise. This enables rapid and training-free segmentation of microstructures across a wide range of alloy systems. The method has been effectively applied to microstructure recognition across a range of alloy systems, including spheroidite identification in high-carbon steels, carbide detection in hot-stamped steels, and phase segmentation in nickel-based superalloys. Notably, without the need for additional training, the segmentation accuracy achieved by this method is comparable to that of supervised models trained on annotated datasets for task-specific applications. Moreover, the model demonstrates superior robustness across varying data volumes and complex alloy microstructures, outperforming supervised learning approaches while significantly reducing the dependence on labeled data for microstructure recognition tasks. The corresponding study was published in Scripta Materialia under the title “Alloy microstructure segmentation through SAM and domain knowledge without extra training.”
2. Microstructure Prediction and Design with Generative Models
Microstructure serves as the pivotal link in the “composition/process–microstructure–property” relationship of multiphase alloys. Although mainstream AI-driven alloy design frameworks incorporate certain quantitative microstructural descriptors, these often fall short in capturing the topological complexity of real microstructures. As a result, constructing reliable links from composition and processing to complex, real-world microstructure images and their associated properties remains a significant challenge. To address this challenge, Professor Xu Wei’s team proposed a deep generative model-based alloy design framework, termed VAE-DLM, which centers on real microstructure images. The framework integrates a Variational Autoencoder (VAE) with a Multi-Layer Perceptron (MLP): the VAE encodes real microstructure images into a latent space and reconstructs them from the latent vectors, while the MLP directly predicts alloy composition, processing parameters, and material properties from these latent representations. Building on the model’s predictive accuracy, the framework was further integrated with targeted latent space sampling techniques to develop an optimization strategy for the composition, processing, and microstructure of Unified Dual-Phase (UniDP) steels. Experimental validation demonstrated that the UniDP steels designed using this strategy meet the performance criteria of DP780, DP980, and DP1180, while incurring lower costs than existing commercial alloys. More importantly, the framework supports direct prediction of microstructure images for proposed alloy designs, thereby significantly enhancing the interpretability and rationality of AI-assisted materials development. The microstructure images predicted through computational design showed a high degree of agreement with those obtained through experimental validation. Furthermore, the study demonstrates that incorporating realistic microstructure images to complete the full “composition/process–microstructure–property” pathway significantly improves the model’s generalization performance. Comparative analysis with machine learning approaches that do not incorporate microstructural information further confirms the practicality and robustness of the VAE-DLM framework. These findings provide a novel perspective for the design of other multiphase alloy systems. The corresponding study was published in Acta Materialia under the title “Structure-to-Process Modeling Drives Experimentally Validated Unified Dual-Phase Steel.”
3. Mechanical Response Mechanism Analysis with Multimodal Deep Learning
The generalized analysis and in-depth understanding of mechanical response mechanisms—such as work hardening in metallic structural materials—have long been key research focuses in the field. While traditional mechanical modeling approaches based on crystal plasticity theory have solid physical foundations, in practical applications they require complex constitutive equations and parameter calibrations tailored to different compositions and processing conditions in order to accurately capture the intrinsic microstructure–mechanical property relationships. As a result, the need for complex constitutive equations and parameter calibration significantly constrains the generalizability and efficiency of such models. To address these challenges, Professor Xu Wei’s team proposed a novel framework—CP-CNN (Crystal Plasticity-guided Convolutional Neural Network)—in which crystal plasticity theory guides the application of multimodal deep learning to analyze mechanical response mechanisms. Specifically, the model introduces dot-product-based coupling to deeply integrate numerical and image-based modalities, allowing composition, processing parameters, and local stress distributions to interact coherently under physically informed constraints. The resulting fused feature matrix is fed into a convolutional neural network, allowing for accurate prediction of stress–strain curves, work hardening behavior, and the onset of necking in dual-phase steels, while preserving physical interpretability. The model achieves an average prediction accuracy of 96.6%. Notably, it also captures stress concentration features at phase boundaries during plastic deformation. Without the need for tuning constitutive parameters, the model consistently identifies deformation mechanisms across various alloy systems. This demonstrates its robust capability to characterize microscale mechanical behavior and provides a solid foundation for advancing generalized studies on material deformation mechanisms. The related work was published in Acta Materialia under the title “Fitting-free mechanical response prediction in dual-phase steels by crystal plasticity theory guided deep learning.”
4. Transfer Learning Guided by Microstructure Evolution Mechanisms for Predicting and Designing Long-Term Performance
The long-term performance of metallic structural materials is closely governed by their microstructural evolution—for example, the relationship between precipitate coarsening and creep resistance in heat-resistant steels. To enable reliable prediction of creep life and efficient alloy design, Professor Xu Wei’s team developed an uncertainty-aware deep learning framework named PM-TR-BCNN, which integrates microstructure evolution mechanisms with transfer learning strategies. The framework captures the interdependence between short-term tensile strength and long-term creep behavior by incorporating physically grounded strengthening factors related to precipitate coarsening. These factors are jointly modeled within a Bayesian neural network to predict both creep life and associated uncertainty. Leveraging the model’s predictive accuracy, the study integrates a multi-objective genetic algorithm to formulate an alloy design strategy that simultaneously optimizes creep performance and minimizes prediction uncertainty. Under conditions of 650 °C and 140 MPa, the proposed strategy enabled the development of three novel creep-resistant steels. The best-performing alloy achieved a creep life more than twice that of existing commercial materials. Experimental validation showed excellent agreement between predicted and actual creep lifetimes, confirming the framework’s robustness and practical utility. Moreover, the study demonstrates that integrating physics-informed metallurgical mechanisms—such as precipitation strengthening—with small-sample transfer learning significantly enhances the generalization capability of the model. Furthermore, the work systematically underscores the critical role of uncertainty modeling in the alloy design process. These findings present a new paradigm for the accelerated and reliable design of high-temperature alloys. The research was published in Acta Materialia under the title “From creep-life prediction to ultra-creep-resistant steel design: An uncertainty-informed machine learning approach.”
The series of studies conducted by Professor Xu Wei’s team this year demonstrates the transformative potential of AI-driven microstructural analysis in advancing materials science and highlights its central role in fostering innovation across the field. Continued advances in specialized, adaptable AI architectures—designed for microstructure recognition, prediction, and deep information mining—hold promise for fundamentally reshaping core areas of research, including property prediction, mechanistic understanding, and alloy design.
