Raman-Grounded Deep Learning for Calcium Carbonate Polymorph Sensing during Microfluidic Biomineralization

Microbially induced calcium carbonate precipitation (MICP) in microfluidic systems produces spatially and temporally heterogeneous precipitates whose polymorph composition governs material outcomes but is difficult to monitor non-destructively in situ. We present a Raman-grounded deep learning framework that decouples spectroscopic labeling from optical inference: each Raman spectrum is matched to an individually segmented crystal instance, and a five-channel EfficientNet-B0 classifier trained on these per-instance labels assigns polymorph identity to the remaining crystals from brightfield and polarized light alone. Trained on 550 Raman-labeled instances and applied across 5123 segmented crystals spanning five experimental datasets, the framework achieved 97.3% test accuracy across three polymorph classes-calcite, early-stage vaterite, and late-stage vaterite-at individual-crystal resolution. Late-stage vaterite emerged as a third optically distinct class, identified from 48 h imaging and confirmed by Raman to retain vaterite phase identity despite its altered optical signature. The framework resolved systematic effects of solution composition, bacterial density, pH, and time on polymorph composition and captured the dissolution of vaterite toward a calcite-dominated state within a single device. Operating at individual-crystal resolution across heterogeneous crystal populations without requiring droplet confinement, the approach provides a practical route to non-destructive polymorph monitoring in MICP and living building materials applications.