Learnable Adaptive and Robust Controller for a Two Particle Carbonate Precipitation Process

In this work, we introduce a novel learning-based controller tailored for autonomous control of a batch-type precipitation process involving calcium and magnesium carbonates. The process takes in fluid containing valuable materials such as Ca
⁺²
and Mg+2 ions, along with impurities and seed particles, to facilitate the sequential precipitation of these ions into their respective carbonates. The controller's goal is to attain a specified size of the precipitated particles under different process uncertainties. Here the residence time, i.e. the time allowed for the ions to remain in fluid phase, is used as the manipulation variable. The controller is designed as a solution to a stochastic optimal control problem and implemented using machine learning techniques. For the prediction model, we use convolutional neural networks (CNN) and for the control synthesis, we use a type of recurrent neural networks (RNNs). The designed control is learnable, adaptable to varying process dynamics and robust to random disturbances in the process, thus resulting in a learnable adaptive, and robust controller (LARC). The effectiveness of LARC is validated through different simulation-based tests.