Modeling And Optimization Of Fermentation Process Based On Machine Learning

Industrial biotechnology is one of the cores to achieve sustainable industrial development.As an important component of industrial biotechnology,fermentation engineering utilizes the specific traits of natural or directionally modified microorganisms and uses modern engineering technology to produce products needed by humans.The operating conditions in industrial fermentation can adjust the growth and metabolic environment of microorganisms,and then improve the efficiency of fermentation production,so the optimization of operating conditions is very important.An accurate and effective mathematical model can not only quantitatively reveal the correlation between the fermentation process variables,and realize the prediction of variables that are difficult to monitor in real time,but also be a prerequisite for further automatic control and optimization.At present,commonly mechanisms or data-driven models are mostly used to describe the external characteristics of the fermentation process,and seldom involve metabolic mechanisms.The metabolic flux regulation model can quantitatively describe the relationship between the operating conditions and metabolic flux in the fermentation process from the perspective of metabolic mechanism,which can provide metabolic guidance for the optimization of operating conditions.Based on the analysis of the process feature and data characteristics,this dissertation employs metabolic flux analysis tools and machine learning methods to propose a new fermentation process modeling and optimization strategy for the fermentation process.The specific research content is listed as follows:(1)Introduce Stacked Denoising Autoencoder(SDAE)into the regression modeling of fermentation process.Based on the cellular metabolic network model and the time-varying data of the extracellular parameters,the commonly used dynamic metabolic flux analysis(DMFA)method is to calculate the dynamic metabolic flux.Some extracellular time-varying parameters need to be tested offline,thereby reducing the real-time nature of DMFA modeling.In view of the process characteristics of the fermentation process,the multi-sampling rate among different variables and the characteristics of measurement noise,the paper uses the semi-supervised learning method SDAE to predict the required extracellular time-varying parameters in real time.First,use the unlabeled sample data in the process to carry out unsupervised pre-training of SDAE layer by layer,and then use the labeled data to perform supervised fine-tuning of the overall network.(2)Based on SDAE’s effective prediction of extracellular time-varying parameters,analyze firstly the metabolic information in the cellular metabolic network,and establish the dynamic flux equilibrium equation based on the intracellular pseudo-steady state hypothesis and the principle of mass conservation.According to the metabolic flux linear changes between adjacent DMFA moments,combined with time-varying data of extracellular parameters,solve the equilibrium equation to estimate the dynamic metabolic flux.By reasonably selecting the operating condition variables and the corresponding dynamic metabolic flux as the modeling sample data set,the metabolic flux regulation model can be constructed.(3)Optimize the fermentation operating conditions based on the metabolic flux regulation model of SDAE-DMFA.The optimization strategy based on the metabolic flux regulation model is designed,that is,the metabolic flux of product synthesis,cell growth and substrate consumption reaction are taken as the optimization goal,and the constraint conditions such as model constraints and upper and lower limits of decision variables are analyzed.The fast non-dominated sorting genetic algorithm(NSGA-II)is used to solve the multi-objective optimization problem.The acceleration rate of substrate flow and cold water flow in the process of penicillin fermentation are optimized to effectively reduce substrate consumption and increase yield.