Liquid-Liquid Dispersion Performance Prediction and Uncertainty Quantification Using Recurrent Neural Networks.

Area of Science:

• Chemical Engineering
• Computational Science
• Artificial Intelligence

Background:

• Complex multiphase systems, such as stirred and static mixers, require accurate performance predictions.
• Traditional methods may struggle with the multivariate and multistep time-series nature of mixer performance data.
• Computational fluid dynamics (CFD) simulations offer high-fidelity data but can be computationally intensive.

Purpose of the Study:

• To apply recurrent neural networks (RNNs) for multistep and multivariate time-series performance predictions in mixers.
• To compare the efficacy of different RNN architectures (fully connected vs. encoder-decoder) for capturing complex system dynamics.
• To develop a generic workflow for applying RNNs to industrial time-series data.

Main Methods:

• Training RNNs, including those with long short-term memory (LSTM) and gated recurrent unit (GRU) cells, on 3D CFD simulation data.
• Utilizing data on drop size distributions and interfacial areas as a function of physicochemical properties, mixer geometry, and operating conditions.
• Implementing data preprocessing, model exploration, performance visualization, and an ensemble-based uncertainty quantification procedure.

Main Results:

• RNNs can be trained on CFD data to predict mixer performance metrics.
• Encoder-decoder RNN architectures demonstrate a superior ability to learn long-term dynamics compared to simpler RNNs.
• The developed workflow provides a robust method for uncertainty measurement and is adaptable to other industrial applications.

Conclusions:

• RNNs are effective tools for predicting the performance of complex multiphase systems like mixers.
• The choice of RNN architecture significantly impacts the model's capability to capture underlying system dynamics.
• The presented generic workflow facilitates the application of machine learning for time-series analysis in diverse industrial contexts.