A machine learning model guided by physical principles for biofilter performance prediction

Abstract

Abstract Despite the critical role of biofilters in water quality and sustainability, predicting their performance remains challenging due to the complexity of microbial interactions and limitations of sparse, high-dimensional datasets. Here, we introduce EnviroPiNet, a novel physics-guided AI framework designed to predict biofilter performance by accurately modeling carbon concentration dynamics. EnviroPiNet incorporates a physics-inspired backbone that enables the model to learn the physical properties of complex environments, ensuring predictions are grounded in system behavior. Additionally, we implement an ensemble hybrid approach to identify and extract key parameters essential for accurate carbon concentration predictions. We benchmark EnviroPiNet against conventional methods that lack physics-guided variable selection, demonstrating its superiority in identifying variables critical to biofilter performance evaluation. Trained on biofilter datasets, EnviroPiNet achieves a high coefficient of determination ( $$\text {R}^{2}$$ = 0.9) on test sets, highlighting its predictive accuracy and robustness.