| Texto completo | |
| Autor(es): |
Toda, Daniel H. R.
;
Bressane, Adriano
;
Negri, Rogerio G.
;
Formiga, Jorge K. S.
;
dos Reis, Adriano G.
;
Bankole, Abayomi O.
;
Bankole, Afolashade R.
;
Sharifi, Soroosh
;
Moruzzi, Rodrigo
Número total de Autores: 9
|
| Tipo de documento: | Artigo Científico |
| Fonte: | Chemical Engineering Science; v. 321, p. 13-pg., 2026-02-01. |
| Resumo | |
Accurate prediction of settling velocity, a critical characteristic of multiphase flow, is vital for optimising sedimentation units in water treatment. Traditional models, such as Stokes' Law, often fail to capture the complex behavior of fractal aggregates, whose irregular structures influence drag coefficients and settling dynamics. Studies have shown that these aggregates may settle faster or slower than perfect spheres, emphasizing the critical role of aggregate morphology beyond simple size parameters. While machine learning (ML) models can address non-linearities, they often lack physical grounding, which is crucial for accurately predicting the settling velocity of fractal aggregates influenced by fluid-particle interactions and drag dynamics. To address this, we developed and compared two predictive models: a traditional Gradient Boosting Regressor (GBR) and a GBR extended through physics-based feature engineering (PFE GBR). The latter incorporates features derived from Stokes' Law, Reynolds number, and drag force, enhancing its alignment with sedimentation theory. This work establishes a critical baseline, demonstrating the efficacy of integrating physical principles into a widely used and robust machine learning algorithm like GBR. Both models were rigorously evaluated using five-fold cross-validation. The PFE GBR outperformed the traditional model, achieving a test R2 of 0.951 versus 0.887, and reducing the mean absolute error from 234.31 mu m/s to 136.05 mu m/s, alongside a 30.1 % reduction in RMSE. These improvements demonstrate that embedding physical principles via feature engineering into ML frameworks enhances predictive accuracy while implicitly guiding model generalization and interpretability. Our findings underscore the value of PFE approaches in sedimentation modelling, providing actionable insights for improving the design and operational efficiency of water treatment processes. (AU) | |
| Processo FAPESP: | 23/08052-1 - Misturadores de fractal e velocidade terminal dos agregados formados |
| Beneficiário: | Rodrigo Braga Moruzzi |
| Modalidade de apoio: | Auxílio à Pesquisa - Regular |
| Processo FAPESP: | 24/01610-1 - Abordagens baseadas em redes neurais profundas para detecção de mudanças via séries de imagens de sensoriamento remoto |
| Beneficiário: | Rogério Galante Negri |
| Modalidade de apoio: | Auxílio à Pesquisa - Regular |