AI-Assisted Optical Microstructural Characterization of Mineral Insulation Materials: Toward Predictive Thermal Performance Modeling

The performance of mineral-based insulation materials is governed by their microstructural configuration, which directly influences heat transfer mechanisms across multiple spatial scales. This study presents an integrated methodology combining optical microscopy and Artificial Intelligence (AI) to characterize and compare five widely used mineral insulation materials: perlite, glass wool, rock wool, calcium silicate, and foam glass. High-resolution optical images are processed through distinct AI models to extract a comprehensive set of geometric (e.g., pore diameter, pore size distribution, fiber dimensions) and morphological (e.g., solid matrix fraction, porosity, density, pore connectivity, anisotropy, tortuosity, surface roughness) descriptors. In addition, the presence and typology of defects are identified and correlated with structural irregularities. A central contribution of the work lies in the estimation of thermal conductivity (λ, W/mK) derived from AI-interpreted microstructural features. The comparative framework critically examines similarities and discrepancies among the outputs generated by the different AI models. The results reveal both coherent trends across materials, such as the relationship between pore connectivity and estimated thermal conductivity, and marked divergences in the interpretation of complex microstructures, particularly in fibrous and highly heterogeneous systems such as glass and rock wool. The comparison highlights differences in model sensitivity to image noise, segmentation robustness, and the capability to capture anisotropy and tortuosity. These variations directly affect the reliability of derived parameters and, consequently, the estimation of the λ value. The study highlights strengths and limitations of current AI-based diagnostic approaches, identifying conditions under which automated analysis is reliable and where expert interpretation remains necessary. Thus, it frames AI-based analysis as a spectrum of model-dependent interpretations requiring critical validation with laboratory tests and on-site analyses. Within the framework of sustainable building practices, the proposed methodology supports the development and selection of high-performance, low-impact insulation materials by providing deeper insight into structure–property relationships, while simultaneously assessing the reliability of digital diagnostic tools. In this sense, this research contributes to advancing data-driven material optimization, informing energy-efficient design strategies, and clarifying the practical role of AI in supporting construction processes.