24 April 2026 in SSRN Electronic Journal
Scanning electron microscopy (SEM) during thermal-shock durability testing of thermal barrier coatings (TBCs) is limited to a few checkpoints, which limits microstructural evolution analysis due to sparse imaging data. Here we present FracFlowFormer, a data-efficient forecasting framework that learns from limited early-stage SEM observations and predicts late-stage fracture and pore evolution in TBCs. To address the issue of coarse temporal sampling, we reconstruct more densely sampled temporal sequences employing TV-L1 dense optical-flow interpolation, thereby generating intermediate states that maintain fracture- and pore-relevant morphology. FracFlowFormer forecasts subsequent microstructural evolution using a hybrid architecture that combines a U-Net encoder–decoder and a spatio-temporal fusion transformer for long-range temporal SEM image generation, enabling high-fidelity predictions that capture microstructural evolution in TBCs during thermal-shock cycling, with 89.95% SSIM, 0.049 LPIPS, and 23.18 dB PSNR. An optical-flow branch provides a microstructural evolution cue that complements grayscale appearance and temporal aspects information. The framework reconstructs high-fidelity, time-resolved image sequences from sparse early-stage observations and demonstrates descriptor-based characterization of microstructural evolution, thereby transforming TBC characterization from static inspection into evolution-aware analysis of degradation and failure.
20 June 2026 in Materials & Design