New research project in the scope of DFG SPP2364 "Autonome Prozesse in der Partikeltechnik"
The aim of this project is to develop the individual process steps for CO2 storage and the production of recyclable carbonates and to link them to form an autonomous process chain. The autonomy of the process is realized via a Self-Learning Robust Autonomous Control(ler) (SLARC) algorithm. The SLARC is an adaptive algorithm, which consists of three interconnected arametric statistical estimators. The estimators are primarily meant to be implemented via (deep) neural networks (DNN), thus can be trained using data and process knowledge (model). Thus, the SLARC controller is to be understood as a kind of hybrid model which combines physical laws with information gathered from data. The former makes it possible to detect unfavorable/unphysical states by constraining to physics of the process while the latter makes it possible to detect changes in particle properties during operation and to adapt to the current state of the process. These two combined with the reinforced online learning feature of the SLARC makes it a self-learning and autonomous algorithm. Additionally, since SLARC is built using a stochastic optimization formulation, it is also designed to be robust to random disturbances. Altogether, the objective of the SLARC is to autonomously keep the process near the optimum operating point by intelligently adapting the (intermediate) target set-points and control variables.
The optimum operating point of the process is the state of highest possible ion concentration of Ca2+ and Mg2+ combined with a high filtrate flow for the filtration step. More specifically, for the selective precipitation, the objective of the process and the control system is to achieve the highest possible product purity while simultaneously maintaining the desired particle size and high solids mass. The observed variables for the selective precipitation process are limited to easily measurable quantities such as mass flows or the particle size distribution from which, in turn, further quantities such as the composition and also kernel functions are derived. Here, the autonomous and self-learning feature of SLARC must be able to estimate growth rates and kernel function parameters which are varying due to varying input composition.