Modelling and parameter identification of penicillin fermentation using physics-informed neural networks

Abstract

With the rapid development of machine learning technology and computer science, artificial neural networks have become an effective and popular method in the existing modelling research of penicillin fermentation process. Although these networks can capture the complexity of the fermentation process, they may lead to overfitting and require large amounts of data. In addition, the inference of the model on the data may not satisfy the physical laws. In this paper, a penicillin fermentation modelling method based on physics-informed neural networks is proposed. The fermentation mechanism equations are combined with the neural networks to develop the model as constraints. First, a general penicillin fermentation mechanism model is built according to known prior knowledge, and then its unknown nonlinear dynamic parameters are identified by physics-informed neural networks. Finally, the successfully trained model exhibits a high prediction accuracy, which not only satisfies the physical laws in the loss function, but also verifies the effectiveness of the proposed mechanism model.