Numerical simulation of surfactant-laden emulsion formation in an un-baffled stirred vessel

Liquid–Liquid Dispersion Performance Prediction and Uncertainty Quantification Using Recurrent Neural Networks

“…Specifically, two network architectures, fully connected (FC) and encoder–decoder, with two different RNN cells, LSTM and GRU, have been finely trained and deployed to predict interfacial area growth, drop count, and their size distribution for future time-steps. The RNN models developed herein have been trained with data obtained from high-fidelity, three-dimensional, CFD simulations of two mixing systems, carried out for stirred , and static , mixers. In particular, the RNN model has been assigned the task of learning and capturing the influence of varying physicochemical properties, operational conditions, and mixer geometrical characteristics on dispersion performance.…”

“…Further details into the equations governing surfactant transport and the specifics behind the interfacetracking algorithm (i.e., LCRM) are found in previous publications. 29,31,32,45 Data Preprocessing: Scaling, Smoothing, and DSD Binning. The initial set of features considered in this study consists of two scalar quantities, namely, IA and ND, and a variable-sized list of scalars (i.e., drop volumes) representing the DSD.…”

“…In this work, M was initially set to 20 and 12 for the stirred and static mixing cases, respectively, aiming to provide sufficient resolution for the DSD data based on statistical analyses conducted in previous publications. 29,31 To eliminate outliers and uninteresting features (i.e., bins that mostly remain as 0 throughout the time domain), M was ultimately cut to 10 bins for both mixers (B 0 − B 9 ), dropping the boundary bins at each end of the size range proportionally, yielding 12 features overall for the RNN to handle. Finally, the ND and IA features were scaled and smoothed individually to mitigate potential biases arising from their substantially different scales [∼O(10 2 ) vs ∼O(1)] and the sharply fluctuating trend observed for the ND data.…”

“…These parameters are used to label the surfactant-laden cases employed herein and broadly describe the nature of each surfactant modeled (e.g., highly/weakly adsorptive/desorptive, etc.). Further details into the equations governing surfactant transport and the specifics behind the interface-tracking algorithm (i.e., LCRM) are found in previous publications. ,,, …”

“…Inspired by the work reviewed before, this study seeks to develop an inexpensive time-series model using RNNs by capitalizing on a comprehensive set of high-fidelity three-dimensional CFD simulations, some of which have been exploited in recent works − to unravel the fundamental governing mechanisms underlying extensively utilized mixing systems handling L–L dispersions across a range of industrially relevant scenarios. These simulations have been conducted with a state-of-the-art DNS code, which comprises a hybrid front-tracking/level-set interface-tracking algorithm, embedded along a well-validated multiphase solver for surfactant transport at the interface and in the bulk phase .…”