Author(s): Loes Verhaeghe; Jan Verwaeren; Sina Borzooei; Benedek G. Plosz; Elena Torfs

Linked Author(s):

Keywords: Aeration; Computer vision; Hybrid models; Microbial images; Sedimentation

Abstract: Microbial images of activated sludge are often available at Wastewater Resource Recovery Facilities (WRRFs), yet they have not been integrated into mechanistic models. This study investigates how image-derived microbial information can be integrated into Hybrid Models (HMs) to enhance predictions of aeration efficiency and sludge settling behaviour. We explored two integration strategies: predicting residuals of mechanistic models (parallel HM) and predicting mechanistic model parameters (serial HM). The serial configuration linking image features to intrinsic process parameters showed superior predictive stability. Four data-driven methods were compared to process the images, with a pretrained convolutional neural network showing the greatest potential. The models were applied to two datasets from different WRRFs to assess the influence of dataset characteristics on predictive performance. The results indicate that predictive performance is strongest when training data includes meaningful variation driven by microbial dynamics rather than operational changes. Overall, this study demonstrates the potential of leveraging microbial images in HMs, particularly for predicting mechanistic model parameters that reflect intrinsic microbial-driven processes, such as settling and aeration.

DOI:

Year: 2026