Modeling And Optimization Of Continuous Crystallization Process With Physics Informed Neural Network

Solution crystallization is a widely used process for the separation and purification of substances in the chemical and pharmaceutical industries.Compared to the traditional batch crystallization,continuous crystallization offers advantages in terms of crystal size distribution consistency and better process repeatability.Continuous crystallization with tubetype crystallizer is becoming widely accepted by the industry as a mainstream crystallization method.How to model and optimize the continuous crystallization process has also become a topic worthy of in-depth study.In this thesis,Continuous Oscillatory Baffled Crystallizer(COBC)is used as the crystallizer.The online estimation of crystal size distribution with traditional numerical algorithm is time-consuming and the value needs to be recalculated when the initial value drifts.To address the above problem,combining the mechanism model of crystallization dynamics with deep learning model,a numerical model based on the Physics-Informed Neural Network(PINN)is proposed for the solution and optimization of the continuous crystallization process model.The main work is as follows:(1)Building a crystallization process model based on PINN: study the mechanism model of the continuous crystallization process and combine it with the COBC structure to build a Population Balance Model(PBM)that controls the dynamic behavior of the crystallization process.This model is essentially a nonlinear Partial Integral Differential Equation.Because the method of solving this equation is complicated.PINN is introduced to solve it.According to the characteristics of deep learning,the Population Balance Equation(PBE)is transformed for Deep Neural Network(DNN)training,and the output of the network is the evolution process of Crystal Size Distribution(CSD).(2)Using the PINN model to design the crystallization process: construct a Physics model based on the modified PBE(MPBE)and other constraints,and for the long-term continuous crystallization process,the initial value will drift,and the above algorithm is modified to solve the model.Secondly,the crystal size distribution curve solved by PINN and the desired CSD are minimized to obtain the optimal constant super saturation,so as to calculate the cooling temperature curve of the solution.(3)Taking glutamic acid as the crystallization object,and simulate the continuous crystallization process according to the introduced PINN: by establishing the CSD change process under the same initial distribution and different constant super saturation,and the CSD change under different initial distributions and the same super saturation.The model of the process can then be used to obtain the temperature curve of the solution in different states.Using the solved CSD to design the temperature curve to get the desired CSD.The simulation results show that under the optimal constant super saturation,the CSD of the modified PBE is similar to the target CSD of the PINN approximation,which verifies the effectiveness of the model for the crystallization process.