Distribution of concentration in hybrid computation and finite element method in assessment of the flow properties in the two-phase contactor

This research paper explores the application of several machine learning (ML) models for simulating ozone content in a hybrid membrane ozonation (OZ) process. The data for ML computations were extracted from computational fluid dynamics (CFD) analysis of the membrane system. The dataset comprises over 10,000 data points derived from CFD simulation for the feed side of a two-phase membrane contactor. The ML models employed in this research include Multi-Layer Perceptron (MLP), Gated Recurrent Unit (GRU), Decision Tree Regression (DTR), and Huber Regression (HBR). To optimize the performance of these models, Bayesian Hyper-parameter Optimization (BHO) was adopted for hyper-parameter tuning. The results indicate that MLP achieved an R
2
of 0.99523, root mean square error (RMSE) of 0.07995, and maximum error of 0.45358, revealing strong accuracy and consistency. GRU attained the highest R
2
of 0.99611 with RMSE of 0.07282 but with slightly higher variability. DTR performed well with an R
2
of 0.99305 and RMSE of 0.09738, while HBR, with an R
2
of 0.85574 and RMSE of 0.39933, demonstrated weaker performance. These findings revealed MLP as a robust model for predicting ozone concentration in a hybrid membrane-OZ process to improve the efficiency of the process by uniform distribution of ozone throughout the feed solution, which consequently improves the degradation/separation of water pollutants.