Numerical simulation and optimization of Lonicerae Japonicae Flos extract spray drying process based on temperature field verification and deep reinforcement learning

This paper presents a computational fluid dynamics (CFD) method combined with deep reinforcement learning to simulate and optimize the spray drying process of Lonicerae Japonicae Flos (LJF) extract. The computational model firstly incorporates the drying kinetics information, which was experimentally determined by drying of individual droplets. Secondly, the difference between this study and previous work is that a distributed optical fiber temperature measurement system (DTS) was used to measure the temperature field of a pilot-scale drying tower for model verification. The mean percentage errors between the experimentally measured temperature and the simulated values at 3 heights (0.18 m, 0.48 m, and 0.78 m) were 8.8%, 7.1%, and 3.1%, respectively. The measured temperature in the drying tower is consistent with the simulation, which can well explain the change of droplets during the drying process. Based on experimental and simulation data, a powder yield prediction model was established. Deep reinforcement learning model was then applied to continuously interact and iterate with the prediction model, realizing the automatic optimization of the spray drying process. The results show that the process can be optimized to increase the powder yield by around 5%. The model can thus be used as a basis for equipment improvement and to provide optimal operating conditions for spray drying process to replace traditional empirical adjustment method.