An interpretable grid-resolution machine learning framework for predicting biofilm detachment

Biofilm detachment affects biomass release, microbial dispersal, and operational stability in water systems. However, predicting where and when detachment occurs remains difficult. Most existing models rely on bulk scale descriptions and do not resolve microscale structural heterogeneity. Here, we developed a framework at grid resolution that integrates in situ confocal laser scanning microscopy, computational fluid dynamics, and interpretable machine learning to predict localized biofilm detachment from coupled structural and hydrodynamic information. Three-dimensional Shewanella oneidensis MR-1 biofilms were discretized into micrometer scale grids and paired with CFD derived local shear fields. This workflow produced a dataset of 26,653 local observations linking biofilm morphology, hydrodynamic exposure, and detachment response. Among ten regression models, the Extra-Trees Regressor showed the best overall predictive performance and robustness and was therefore used consistently for model interpretation, grouped validation, external validation, and inverse prediction. Model interpretation indicated that local detachment predictions were mainly associated with the balance between structural vulnerability and attachment support. Greater local thickness was associated with higher predicted detachment, whereas the basal layer showed a stabilizing association. The thickness and shear-rate transition ranges identified by SHAP, PDP, and ICE analyses should be interpreted as system specific, data driven model response regions rather than universal mechanistic thresholds. External validation in controlled porous media and simulated drinking water pipe systems supported preliminary laboratory scale transferability, but applicability domain analysis showed that the external predictions contained an extrapolative component.