Random Forest and Autoencoder Data-Driven Models for Prediction of Dispersed-Phase Holdup and Drop Size in Rotating Disc Contactors

Abstract

Linear regression models are traditionally used to capture the relation between the input and output variables. Linear models cannot account for the nonlinear relations in the data. Hence, the prediction models may not be accurate. For this reason, machine learning-based models are being increasingly used. For modeling, design, and scaleup of rotating disc contactors (RDCs), rational estimation of dispersed-phase holdup and drop size is crucial. We have employed random forest (RF) and autoencoder−RF-based models for the prediction of dispersed-phase holdup and drop size in RDCs. Our results show that both these models predict drop size quite well. For holdup, the autoencoder−RF combination predictions are not satisfactory. The standalone RF model predictions generalize very well. RF-based models can be further used for prediction of different variables of interest in RDCs.

Interpretation of Drop Size Predictions from a Random Forest Model Using Local Interpretable Model-Agnostic Explanations (LIME) in a Rotating Disc Contactor

“…Since, we have not generated any new data in our work, we have adopted the definition of the drop size as per the available literature. More comprehensive information on the experimental data and the empirical correlations is made available as Supporting Information in our previous work 2 and is not reproduced here for the sake of brevity. The Supporting Information can be accessed at https://pubs.acs.org/doi/10.1021/acs.iecr.0c04149…”

“…The empirical correlations and the AARE values can be found in the Supporting Information of our previous work. 2 The random forest model with a 5-fold cross-validation was developed as part of our previous work. 2 Machine learning models have been widely employed in chemical engineering for estimation of bubble size and holdup in a bubble column, flow regime identification, estimation of mass transport coefficient, etc.…”

“…However, it was observed that the linear regression model gave a poor prediction performance as R 2 and AARE for the test set were found to be 0.6230 and 27.21%, respectively. 2 The random forest model based on random forest models with top features was developed. The prediction performance of the random forest for a test drop size data set was found to be R 2 = 0.8725 and AARE = 15.7946%.…”

Review of Machine Learning for Hydrodynamics, Transport, and Reactions in Multiphase Flows and Reactors

“…Persistent lack of physical comprehension continuously stymies preferable prediction performance of the key parameters in multiphase flow and reactor systems, although scientists have made systematic contributions to experimentally formulated correlations throughout the past decades. − The correlations of the key parameters in multiphase units are commonly expressed by gas/liquid/solid phase properties, operating conditions (e.g., phase concentration, velocity, and temperature), devices configurations (e.g., height and diameter), or a combination of them in dimensionless forms such as Archimedes, Froude, Nusselt, Reynolds, Sherwood, and Weber numbers. However, the prediction discrepancies between the existing empirical correlations of key parameters such as the particle entrainment and minimum fluidization velocity in gas-particle riser flows can reach several orders of magnitude. , Fortunately, the advanced research and development of flexible ML tools have the potential to complement the incomplete knowledge to boost the prediction ability of key multiphase field parameters such as mass flow rate/flux, − minimum fluidization velocity, , mixing rate/index, , overall/local hold-up, − pressure/pressure drop, − velocity, ,− temperature, − and other parameters − in multiphase/particulate flows and reactors.…”

Prediction of the Aqueous Solubility of Compounds Based on Light Gradient Boosting Machines with Molecular Fingerprints and the Cuckoo Search Algorithm

“…Advanced machine learning methods expand the application scope of the QSPR model. − In recent years, ensemble learning methods, especially random forest (RF) and light gradient boosting machine (LightGBM), have yielded satisfactory results in dissolution prediction. , In 2021, Ye et al predicted the solubility of compounds in organic solvents with the LightGBM algorithm, which showed better generalization ability compared to deep learning and other traditional machine learning algorithms.…”