Recognition of fossil keratose sponges in carbonates using petrographic and machine‐learning classification

ABSTRACT Filamentous, anastomosing spar‐filled networks in Phanerozoic carbonates have been interpreted as the mineralised remains of keratose (non‐spiculate) sponges, yet their recognition remains contested because fabrics of broadly similar appearance may arise from microbial, diagenetic or physical processes. This study combines direct fossil–modern petrographic comparison with machine learning classification to evaluate and quantitatively test recognition criteria for these fabrics. Well‐preserved Upper Ordovician filamentous networks from the Xiazhen Formation, South China, are compared with thin sections of modern keratose sponges prepared by identical methods. The modern spongin networks, when embedded and sectioned, reproduce the key features observed in the fossils: dense, three‐dimensional networks of uniformly thick filaments with ordered branching, coherent external margins and a two‐scale void hierarchy consistent with main exhalant and subordinate canals. A machine learning pipeline trained on domain‐specific morphological features extracted from ~700 thin‐section images across multiple published collections tests whether the criteria capture statistically distinct signatures across four fabric categories (keratose, lithistid, clotted and peloid). Distance‐weighted K‐Nearest Neighbours (KNN) with Manhattan distance achieved 82.14% accuracy, outperforming all convolutional neural network (CNN) architectures (best: 73.91%). This suggests that for carbonate‐microfabric classification with limited training data, domain‐specific feature engineering can outperform deep learning. Independent evaluation of published fossil specimens yielded 84.2% sponge‐related classification, and modern sponge thin sections achieved 100% sponge detection accuracy, supporting reliable petrographic identification. This is consistent with a taphonomic pathway in which early automicrite precipitation captures the spongin framework, followed by fibre degradation leaving moulds later filled by sparry calcite. Evaluation of alternative interpretations (clotted micrite, meiofaunal tunnelling, fluid escape, biofilm channels and thermal alteration) indicates that none reproduces the complete suite of diagnostic features of keratose sponges. The integrated petrographic‐quantitative approach provides a reproducible framework that reduces observer subjectivity and enables independent verification of these interpretations across research groups, helping constrain interpretations of reef construction and the role of sponge‐microbial consortia in carbonate facies analysis.