Interpretable Material Spatial Intelligence for Discovery of Governing Microstructural Features

Technical Summary: Interpretable Material Spatial Intelligence for Discovery of Governing Microstructural Features

Problem Statement
Material properties, particularly in metallic systems, are governed by complex spatial and temporal interactions across multiple length scales, ranging from atomic defects to macroscopic heterogeneity. Traditional approaches to linking microstructure to mechanical behavior often rely on handcrafted descriptors or aggregated representations (e.g., average grain size, phase fractions) that overlook the critical role of spatial organization and local deformation heterogeneity. While Machine Learning (ML) is widely used, many existing methods fail to capture the collective behavior of local states or the specific interactions that govern macroscopic properties. Furthermore, conventional physics-based models (e.g., crystal plasticity) often require extensive parameterization and computational resources, while data-driven approaches frequently lack interpretability, making it difficult to identify the fundamental mechanisms controlling property trade-offs, such as the strength-ductility or fatigue-strength dilemmas.

Methodology: Materials Spatial Intelligence (MSI)
The authors introduce Materials Spatial Intelligence (MSI), a framework inspired by spatial intelligence that learns directly from multimodal spatial observations of material systems. The methodology proceeds through the following stages:

1. Data Acquisition and Modality: The study utilizes high-resolution, large-area experimental datasets from ten FCC alloys (including wrought, additively manufactured, and heat-treated variants). Data is captured using Electron Backscatter Diffraction (EBSD) for microstructural states (crystallographic orientation, phase distribution, defects) and High-Resolution Digital Image Correlation (HR-DIC) for deformation states (strain, lattice rotation, deformation event intensity) under simplified loading conditions.
2. Spatial Encoding: Instead of averaging descriptors, the framework treats materials as spatially resolved fields. Large microstructure and deformation maps are partitioned into correlated local regions (tiles). These paired local states are independently encoded into shared latent representations using Convolutional Neural Networks (CNNs). This process preserves spatial relationships and feature correlations without relying on predefined engineered descriptors.
3. Prediction and Trade-off Analysis: The latent representations are fed into a modular regression model to simultaneously predict six mechanical properties: yield strength, ultimate tensile strength, two hardening parameters, ductility, and fatigue strength. By predicting multiple properties, the model investigates property trade-offs (e.g., fatigue ratio vs. yield strength) and identifies local states that break average trade-off behaviors.
4. Interpretability via LIME: To move beyond "black box" predictions, the framework integrates Local Interpretable Model-Agnostic Explanations (LIME). The authors adapt LIME for materials science by defining physically meaningful features (grains, annealing twins, phases, deformation events) rather than superpixels. Through controlled perturbations (e.g., zeroing out lattice rotation within grains or removing annealing twins), the model quantifies the importance of specific features and spatial regions on predicted properties.
5. Synthetic Optimization: The framework employs an iterative strategy to optimize microstructures. By identifying the most detrimental regions via LIME importance scores and iteratively modifying them (e.g., replacing unfavorable lattice rotations with zero values), the authors generate synthetic microstructural states with improved property profiles.

Key Results

• Latent Space Organization: The UMAP projections of the latent space successfully separate materials based on processing routes, grain structures, and deformation mechanisms. The latent space captures heterogeneities such as partially recrystallized regions and intense plasticity localization.
• Property Prediction: The MSI framework accurately predicts macroscopic mechanical properties from local microstructural and deformation states. It successfully learns the well-known fatigue ratio–yield strength trade-off and identifies local states that exceed average performance.
• Feature Importance and Mechanism Discovery:
  • Grain Size and Lattice Rotation: The analysis confirms that large grains with intense lattice rotation (GROD) and high dislocation density are detrimental to fatigue strength. Conversely, refined grain sizes are beneficial.
  • Deformation Localization: Favorable states exhibit limited deformation localization, whereas detrimental states display intense localized deformation.
  • Annealing Twins: The framework identifies that extended annealing twins within large grains have a detrimental effect on fatigue strength, consistent with prior literature.
  • Spatial Distribution: The LIME analysis reveals that properties are governed not just by the magnitude of features but by their spatial organization. For instance, the framework identified specific regions in W_Invar and AM_HIP_GRX810 alloys where processing-induced lattice rotations negatively influenced both yield and fatigue strength.
• Optimization: The iterative synthetic optimization demonstrated the ability to improve fatigue strength by removing detrimental lattice rotations while managing the trade-off with yield strength. This approach outperforms uniform feature removal by exploiting spatial distribution effects.

Significance and Claims
The paper claims that MSI establishes a new paradigm for microstructure-based materials research by shifting from averaged descriptors to spatially resolved representations. Key contributions include:

• Interpretability: Unlike many deep learning approaches, MSI provides a data-driven, physically interpretable identification of governing microstructural features and deformation mechanisms. It bridges the gap between local microstructural states and macroscopic behavior without relying on simplified physics-based approximations.
• Accelerated Discovery: The framework offers substantial time savings for characterizing advanced properties like fatigue performance, which are typically time-consuming to measure. Predictions can be estimated 30- to 100-fold faster than conventional testing.
• Design Capability: MSI enables feature-driven microstructure optimization. By identifying specific spatial distributions of features that control property trade-offs, the framework provides a pathway for designing alloys that overcome traditional performance limitations.
• Generalizability: The approach is designed to operate across multiple material systems, learning both general features (e.g., slip events) and specific feature-property relationships, thereby transforming how microstructure–property relationships are discovered and engineered.

The authors position this work as a foundational step toward applying spatial intelligence to materials science, leveraging interpretable ML systems to accelerate scientific discovery and material design, while acknowledging that future work requires larger, more diverse datasets and the integration of temporal and chemical modalities to fully capture the complexity of evolving physical systems.