Dehghani H., Perrin H., Belouettar-Mathis E., Patzák B., Belouettar S.
Composites Part C Open Access, vol. 17, art. no. 100625, 2025
A challenge associated with the multiscale modeling of highly consolidated composites is the existing contact effects arising from the manufacturing process. In such cases, porosity significantly decreases as we approach the consolidation surfaces, leading to substantial variations in material behavior in those areas. To address this, we propose an unsupervised machine learning approach integrated with micro-computed tomography (μCT) image processing and Asymptotic Homogenization (AH) for accurate and robust consideration of real microstructure as the basis for an upscaling process. This process employs systems of partial differential equations (PDEs) known as cell problems. This work introduces the Aggregated Vertical Projection Clustering (APC) method, which applies K-means clustering to partition the data into k groups based on porosity. We also present a novel porosity-based periodic cell selection strategy, which uses the Halton sequence to select representative volume element (RVE) cells for each cluster. The workflow generates computational meshes of RVE cells for Finite Element (FE) analysis, solves the cell problems required for upscaling, and calculates the effective heat conductivity. Statistical descriptions and representativity analyses demonstrate that the proposed methodology efficiently and accurately computes the effective properties in these challenging cases.
