引入机理动力学的谷氨酸发酵过程数据模型

打开文本图片集

中图分类号：TQ920.1 文献标志码：A 文章编号：1671-0460（2026）01-0134-07

Abstract： Due to the complexityofbiological fermentation processes，itischalenging to obtain high-precisionmodels using mechanistic modeling approaches. Conversely，data-driven modeling methods lack the kinetic mechanisms of the fermentation process，resulting in poor interpretability.As an alternative，hybrid model structures encompass physical equations such as massand energy conservation， whilealso incorporate neural networks orradial basis functions.These hybrid models combine the advantages of both data-driven and mechanistic modeling，and are widely applied in fermentation process modeling.However， thetraditional hybrid modeladopts Euler discretization methodandrelies on high time scale data.Therefore，a hybrid neural network model structure based on mechanistic dynamics was proposed， whichincorporated unknown dynamics into functional parameters.The hybrid model was trained with the objectives of minimizing state deviation and the differential deviation of states.This model not only predicts changes in substance mass concentrations throughout the fermentation processover time series，but also obtains the temporal variation of important process parameters. A case study on the fermentation production using Corynebacterium glutamicum was conducted to compare thetraditional mechanistic model with the proposed hybrid model structure. Experimental results demonstrate significant improvements in prediction accuracyand generalization ability of the hybrid model， validating the feasibility and superiority of the proposed method.

Key words： Hybrid neural network model; Bioprocess; Fermentation; Dynamic modeling; Physical-information model

发酵过程广泛用于化学品、酶、食品和药品等工业生产，利用微生物将底物转化为目标产品，鉴于发酵过程内部复杂机理，其建模是一项具有挑战性的任务[1]。（剩余7435字）