A CFD-ML augmented alternative to residence time for clarification basin scaling and design

Abstract

Particulate matter (PM), while not an emerging contaminant, remains the primary labile substrate for partitioning and transport of emerging and known chemicals and pathogens. As a common unit operation and also green infrastructure, clarification basins are widely implemented to sequester PM as well as PM-partitioned chemicals and pathogens. Despite ubiquitous application for urban drainage, stormwater clarification basin design and optimization lacks robust and efficient design guidance and tools. Current basin design and regulation primarily adopt residence time (RT) as presumptive guidance. This study examines the accuracy and generalizability of RT and nondimensional groups of basin geometric and dynamic similarity (Hazen, Reynolds, Schmidt numbers) to scale clarification basin performance (measured as PM separation and total PM separation). Published data and 160,000 computational fluid dynamics (CFD) simulations of basin PM separation over a wide range of basin configurations, loading conditions, and PM granulometry (particle size distribution [PSD], density) are examined. Based on the CFD database, a novel implementation of machine learning (ML) models: decision tree (DT), random forest (RF), artificial neural networks (ANN), and symbolic regression (SR) are developed and trained as surrogate models for basin PM separation predictions. Study results indicate that: (1) Models based solely on RT are not accurate or generalizable for basin PM separation, with significant differences between CFD and RT models primarily for RT < 200 hr, (2) RT models are agnostic to basin configurations and PM granulometrics and therefore do not reproduce total PM separation, (3) Trained ML models provide high predictive capability, with (R²) above 0.99 and prediction for total PM separation within ±15%. In particular, the SR model distilled from CFD simulations is entirely defined by only two compact algebraic equations (allowing use in a spreadsheet tool). The SR model has a physical basis and indicates PM separation is primarily a function of the Hazen number and basin horizontal and vertical aspect ratios, (4) With common presumptive guidance of 80% for PM separation, a Pareto frontier analysis indicates that the CFD-ML augmented SR model generates significant economic benefit for basin planning/design, and (5) CFD-ML models show that enlarging basin dimensions (increasing RT) to address impaired behavior can result in exponential cost increases, irrespective of land/infrastructure adjacency conflicts. CFD-ML applications can extend to intra-basin retrofits (permeable baffles) to upgrade impaired basins.

Keywords: CFD; Computer aided design; Green water infrastructure; Machine learning; Retention; Water treatment.