Enhancing Phase Clustering in Nanomechanical Property Maps of Multiphase Materials Using Kernel-Averaged Mechanical Mismatch

1. Problem Statement

The characterization of multiphase materials (e.g., composites, alloys, metallic glasses) via high-speed nanoindentation or Atomic Force Microscopy (AFM) generates large datasets of local mechanical properties (Hardness $H$ and Elastic Modulus $E$). While unsupervised machine learning (clustering) is widely used to segment these maps into distinct phases, current methods face significant challenges:

• Low Mechanical Contrast: When phases have similar stiffness or hardness, standard clustering based solely on $(E, H)$ fails to separate them.
• Diffuse Interfaces: Real materials often possess graded interphases or curved boundaries rather than sharp transitions. Standard algorithms struggle to distinguish gradual transitions from noise or true phase regions.
• Lack of Ground Truth: Experimental data lacks "ground truth" phase labels, making it difficult to validate clustering algorithms objectively. Conversely, existing synthetic benchmarks are often too simplistic (e.g., perfectly separable Gaussian clusters), leading to artificially inflated performance metrics.
• Spatial Ignorance: Most clustering approaches treat data points as independent, ignoring the spatial relationships and local mechanical heterogeneity between neighboring measurements.

2. Methodology

A. Synthetic Benchmark Generation

To overcome the lack of ground truth in experiments, the authors developed a fully parametric Python framework to generate controlled, non-trivial synthetic nanomechanical maps. These datasets include:

• Geometries: Five-region quadrants, rectangular composites (sharp vs. graded interfaces), matrix-fiber composites (circular inclusions), and matrix-particle composites (diffuse, random morphologies).
• Parameters: Tunable mechanical contrast ratios ($E_{ratio}, H_{ratio}$), intra-phase Gaussian noise, and variable interphase widths (from sharp to diffuse).
• Ground Truth: Exact phase labels are known for every pixel, enabling rigorous quantitative evaluation of clustering accuracy.

B. The Kernel-Averaged Mechanical Mismatch (KAMM)

Inspired by the Kernel Average Misorientation (KAM) metric used in Electron Backscatter Diffraction (EBSD), the authors introduce KAMM as a neighborhood-informed feature.

• Definition: KAMM quantifies local mechanical heterogeneity by calculating the average absolute mismatch of the modulus-hardness product ($E \times H$) between a point and its neighbors.
  $$KAMM_i = \frac{1}{N_i} \sum_{k \in N_i} |E_i H_i - E_k H_k|$$
• Variants: The framework also includes component-wise mismatches ($KAEM$, $KAPM$) and relative/coupled variations ($KAMM_{ratio}$, $KAMM_{normprod}$).
• Connectivity: Calculated using first-order (4-neighbors) or second-order (8-neighbors) spatial connectivity.
• Concept: High KAMM values indicate proximity to interfaces or interphases (high heterogeneity), while low values indicate homogeneous regions.

C. Clustering Pipeline

The study evaluates four unsupervised algorithms: Agglomerative Clustering, DBSCAN, Gaussian Mixture Models (GMM), and KMeans.

• Feature Space: Comparisons are made between a baseline 2D space $(E, H)$ and an enhanced 3D (or higher) space $(E, H, KAMM)$.
• Validation:
  • Synthetic Data: Pixel-level accuracy (matching predicted clusters to ground truth) and phase-level property error.
  • Experimental Data: Internal validity metrics (Silhouette score, Calinski-Harabasz index, Davies-Bouldin index) and physical interpretability of the resulting phase maps.

3. Key Contributions

1. Novel Feature (KAMM): Introduction of a spatial descriptor that captures local mechanical gradients, effectively bridging the gap between point-wise mechanical data and spatial microstructural context.
2. Robust Benchmarking Framework: Creation of a diverse suite of synthetic datasets with tunable complexity (noise, curvature, diffuse interfaces) to systematically test clustering robustness where ground truth is available.
3. Algorithm-Specific Insights: Demonstration that the efficacy of KAMM is highly dependent on the clustering algorithm used.
4. Workflow Integration: A complete pipeline from data acquisition to the generation of Representative Volume Elements (RVEs) for Finite Element Analysis (FEA).

4. Results

Synthetic Data Findings

• Improved Accuracy: Incorporating KAMM significantly improves pixel-level classification accuracy for KMeans and Agglomerative Clustering, particularly in low-contrast scenarios and systems with graded interfaces.
  • Example: KMeans accuracy increased from 0.689 to 0.823 with KAMM.
• Algorithm Sensitivity:
  • KMeans & Agglomerative: Benefit from the added spatial coherence.
  • GMM: Shows mixed results; benefits in some matrix-particle configurations but is less sensitive in others.
  • DBSCAN: Performance degrades with KAMM. As a density-based method, the added neighborhood descriptors alter the feature space geometry, causing DBSCAN to misinterpret local gradients as noise, thereby fragmenting clusters.
• Noise Robustness: Under high noise conditions, the benefits of KAMM diminish for centroid-based methods (KMeans, GMM), though Agglomerative clustering remains relatively stable.

Experimental Validation (Ni-SiC Composite)

The method was applied to a real Nickel-Silicon Carbide coating dataset.

• Phase Separation: Using only mechanical features ($E, H$), Agglomerative clustering achieved the best internal validity (Silhouette = 0.31).
• Spatial Coherence: Adding first-order KAMM reduced standard internal metrics (e.g., Silhouette dropped to 0.08) due to increased feature correlation, but it significantly improved the spatial coherence of the phase map. It successfully delineated the Ni matrix, SiC particles, and the intermediate interphase more clearly than mechanical features alone.
• Property Recovery: The phase-averaged properties (Hardness/Modulus) recovered with KAMM were closer to reference values than those obtained with mechanical features alone.
• Higher-Order KAMM: Including second-order KAMM descriptors introduced over-parameterization, blurring the distinction between hard and soft phases and shifting property values toward intermediate values.

5. Significance and Implications

• Reliable RVE Generation: The proposed method enables the generation of more accurate Representative Volume Elements (RVEs) for multiscale modeling. By correctly identifying interphases and avoiding misclassification of boundary pixels, the extracted phase-specific properties are more reliable for Finite Element simulations.
• Interphase Detection: KAMM provides a unique capability to detect and visualize diffuse interphases and graded transitions that are invisible to standard $(E, H)$ clustering.
• Methodological Guidance: The study provides critical guidance for materials scientists:
  • Use KMeans or Agglomerative Clustering with first-order KAMM for robust phase segmentation in complex microstructures.
  • Avoid using KAMM with DBSCAN for this specific application.
  • Limit the use of higher-order KAMM to cases requiring fine-scale heterogeneity analysis, as it may obscure global phase contrast.

In conclusion, this work establishes that integrating neighborhood-based mechanical descriptors (KAMM) into the clustering pipeline is a powerful strategy for resolving phase ambiguities in nanomechanical maps, provided the clustering algorithm is compatible with the resulting feature space structure.