Optimal strategies to control particle size and variance in antisolvent crystallization operations using deep RL

Solution crystallization operations have complex dynamics that are typically lumped into two competing processes namely nucleation and growth. Mathematical models can be used to describe these two processes and their effect on the crystal population when subject to variables like temperature, addition of anti-solvent, etc. To ensure that the crystals meet specific performance objectives, the models need to be solved and the control variables need to be optimized. This has largely been done until now using algorithms from dynamic programming or optimal control theory. Recently, however, it has been shown that learning frameworks like Reinforcement Learning can solve large optimization problems efficiently while offering distinct advantages. In this work, we explore the possibility of computing the optimal profiles of a semi-batch crystallizer to control the mean size and variance using four different deep RL algorithms. The performance on one of the tasks is evaluated experimentally on the anti-solvent crystallization of NaCl in a water-ethanol system.