A Two-Stage Physics-Informed Hybrid Machine Learning Framework for Dynamic Pressure Drop Prediction in Industrial Packed Bed Scrubbers

A Two-Stage Physics-Informed Hybrid Machine Learning Framework for Dynamic Pressure Drop Prediction in Industrial Packed Bed Scrubbers

Parisa Aberi Zaker1 Aidaalsadat Entezari2

1) Senior R&D Expert, Sana Insulation Industrial Group,
2) R&D Manager, Sana Insulation Industrial Group,

IOI Code : ACCD-ACZDCC
Publication : 10th International Conference on Applied Researches in Science & Engineering (carseconf.com)
Abstract :
    Predicting dynamic pressure drop (ΔP) in packed bed wet scrubbers under industrial operation remains a challenge due to bed fouling, volatile organic compound deposition, and complex non-linear fluid dynamics. Traditional hydrodynamic models relying solely on first-principles empirical correlations—such as the Ergun equation—lose the accuracy over long-term operations as packing degradation and plasticizer deposition change void fraction (ε) over time. Meanwhile, standard black-box machine learning algorithms lack physical interpretability, risk of data leakage and fail under new operating conditions. In this study, a Two-Stage Physics-Informed Hybrid Machine Learning Framework is developed to model pressure drop dynamics across a 600-hour operational cycle. In Stage 1, a first-principles hydrodynamic engine models coupled transport dynamics, Ergun viscous-inertial head losses, demister pad resistance, and fan operating point equilibrium. In Stage 2, machine learning algorithms trained to learn the residual between physical predictions and actual sensor measurements. To guarantee the model reliability, a non-leaking dataset partition (60% Train, 20% Validation, 20% Test) implemented. Among all evaluated algorithms, the Hybrid Neural Network (HNN) achieved highest performance, improving R2 coefficient on test data from 0.9260 (Physics-Only) to 0.9731. This approach also reduced the Root Mean Square Error (RMSE) by 39.7% (from 52.00 Pa to 31.36 Pa). The proposed approach introduces a digital twin architecture for real-time industrial monitoring and predictive maintenance.

Keywords : packed bed scrubber physics informed machine learning residual learning ergun equation