Beyond Idealized Voids: Linking Pore Morphology to Hotspot Sensitivity via a Statistical Shape Model

Pore morphology in reactive materials significantly influences hotspot formation under shock loading, yet the relationship between morphological details and hotspot sensitivity remains poorly characterized. Existing studies often rely on idealized shapes [1], such as ellipses, or deep neural generative models that have highly sensitive and uninterpretable latent spaces [2], which limits systematic morphology-sensitivity analysis. In this work, we construct a Statistical Shape Model (SSM) over material microstructure images of 134 different pores. The pipeline extracts polygon contours, normalizes each shape via skeleton-based anchor alignment, resamples each contour to 256 geodesically uniform landmarks, and applies PCA, yielding 20 principal components that capture 97.5% of the morphological variance. We interpret the dominant shape modes through their corresponding geometric deformations, linking the latent representation to intuitive features such as curvature, elongation, and asymmetry. We also add rotation as a latent vector since orientation is normalized in the SSM. [image of SSM pipeline] We evaluate the learned shape space to verify that generated pores remain realistic and representative of the observed morphology distribution, using PCA-space coverage and nearest-neighbor comparisons to the ground-truth shapes. We sample novel pore geometries from the learned SSM and augment the latent representation with orientation. To isolate morphological effects, we constrain the pore area to a fixed target of 1,181 pixels with a mean deviation of less than 20 pixels. We use LatentPARC [3], a fast physics-aware recurrent neural network surrogate trained on single-pore collapse simulations of shocked reactive material to predict the resulting spatiotemporally evolving temperature fields for each generated pore. From the predicted temperature evolution, we extract four quantities of interest: hotspot area, average hotspot temperature, and their rates of change, which are direct indicators of material sensitivity. Our preliminary analysis of hotspot-area growth rate using a Random Forest surrogate reveals that the degree to which a pore is oriented vertically relative to the loading direction is the strongest predictor of hotspot dynamics. Among morphology features, variations in lateral elongation and boundary asymmetry provide the strongest additional predictive signal, followed by changes in the pore’s vertical curvature. Notably, lateral elongation and asymmetry are more predictive of hotspot-area growth than vertical curvature, despite the latter accounting for a larger fraction of the overall geometric variation in the dataset. This indicates that the directions of greatest geometric variance in the SSM are not necessarily the directions most relevant to hotspot evolution, and that only a small subset of physically interpretable morphology modes appears to control hotspot-area growth. We are extending this analysis across all four quantities of interest to identify which morphology modes most strongly govern different aspects of hotspot evolution. These findings will offer new insight into how specific, physically interpretable aspects of microstructure drive shock-induced hotspot behavior; insight that neither idealized void models nor uninterpretable neural generative approaches can provide.