A deep reinforcement learning framework for regime-specific decision in the operation of industrial ammonium sulfate crystallization

Real-time prediction of crystal size in industrial ammonium sulfate (AS) crystallization remains a challenge due to opaque slurry conditions, delayed laboratory measurements, variability across nucleation and crystal growth, and a mismatched sampling interval between inputs and output. Therefore, this study develops a reinforcement-learning router for regime-specific decision-making to guide a multi-model prediction framework in estimating the ratio of large AS crystals. The operating domain is segmented into distinct regimes, and a specialized long short-term memory is trained for each segmented regime to learn the characteristics and local dynamics. A Deep Q-Network (DQN) router evaluates sequential inputs to make a decision and select the best submodel by balancing predictive accuracy and operational consistency. The model is verified on the industrial-scale crystallization system. Based on the results, the DQN-router agent demonstrates effective performance across the distinct operatiing regimes using ten actions. The proposed model achieves a coefficient of determination of 0.961, a root mean square error of 1.810, and a mean absolute percentage error of 1.392 % and outperforms a single-prediction model without decision-making and benchmarked selectors. Analysis of router decisions confirms that the learned policy adapts effectively with shifting operating conditions and resolves overlaps between regime boundaries. A constrained optimization analysis of the model predictions identifies an optimal 19-hour operating pattern, consisting of an 11-hour hold-up period and an 8-hour discharge period, which produces an average large-crystal ratio of 88.38 % ± 2.09 %.