Image-based high-dimensional deep learning framework for spatiotemporal modeling, prediction and transfer learning of microbial cultivation processes.

Abstract

Microbial culture processes exhibit complex and nonlinear dynamics, and accurate monitoring and prediction are essential for effective process optimization. Conventional monitoring methods primarily rely on low-dimensional sensor measurements, such as optical density (OD), dissolved oxygen (DO), and pH, which provide useful quantitative information but do not fully capture the physicochemical characteristics of the culture system. In this study, an image-based analytical framework was developed for high-dimensional analysis and time-series prediction during the cultivation of Bacillus subtilis. A controlled imaging system was established to acquire reproducible time-series images, and deep learning models based on convolutional neural networks (CNN) and convolutional long short-term memory (ConvLSTM) networks were constructed. The CNN provided superior pixel-level prediction performance under the original fermentor condition, while transfer learning (TL) enabled adaptation of the pretrained CNN to independent flask cultivation conditions with strong pixel-level reconstruction. Correlation analysis further showed that several image-derived features retained strong associations with OD600 in CNN-predicted images, whereas relationships with Abs450 were less consistent. These findings demonstrate that controlled image-based modeling can complement conventional process monitoring by providing high-dimensional visual information and a transferable framework for microbial cultivation analysis.