Abstract
•Instant determination of biofilm thickness from OCT scan using Deep Neural Network.•Biofilm OCT scan database creation to train CNN.•Trained CNN used to predict biofilm thickness in UF and MD processes with validation loss < 0.008 µm2 with 300 training images.•Development of regression-based neural network model to correlate biofilm growth with hydrodynamics conditions.•Validation of the model to relate instantly pressure drop with biofilm growth with an absolute error < 2 %.
The growth of biofilm inside the filtration channels module is hard to visualize and has a high propensity to tarnish the process performance. Herein, Deep Neural Networks (DNN) are utilized to gauge biofilm thickness and connect its growth with hydrodynamics parameters to establish a control strategy in an artificially intelligent framework. A database of biofilm images is created from the Optical Coherence Tomography (OCT) scans of various ultrafiltration and membrane distillation experiments.A Convolution Neural Network (CNN) is trained to determine the biofilm thickness from the OCT scans. The trained CNN network can instantly predict 2D and 3D biofilm thickness for unseen OCT images of different filtration technologies (ultrafiltration or membrane distillation) with reasonably accurate prediction performance compared to manual calculations. A mean squared error of less than 0.008 µm2 is achieved for a set of 300 testing images while determining the biofilm thickness. Further, a synthetic database is created using a theoretical model to associate the cylindrical channel pressure drop (100–1500 mbar/m) with channel thickness (up to 787 μm) that hypothetically relates growing biofilm with the channel hydrodynamics and geometric parameters (velocity 0.1–0.16 m/s, channel radius 10–21 cm, viscosity 0.0007–0.003 Ns/m2). A Non-Linear Regression-DNN (NLR-DNN) is trained and predicts output quantities (either channel pressure drop or biofilm thickness) below 2 % absolute error against the analytical solution. The validating dataset is compared directly with the theoretical model, and a good fit of R2 = 0.9999 was achieved. The developed framework can potentially be deployed in desalination plants for early decision-making and preventive controls.