Neural network modeling of the dynamic inactivation of probiotics during single droplet drying for improved cell viability

Dairy products containing probiotics are often dried to improve shelf life and facilitate transportation. A reliable dynamic inactivation model has long been pursued to optimize the production by maximizing probiotics' sur-vival during drying. How to take care of the dynamic drying process experienced by the cells for precise pre-diction of their survival remains a challenging task. In this work, a multi-task convolutional self-attention network (CSAN) has been developed for dynamic modeling of probiotics inactivation during single droplet drying (SDD). The convolution self-attention approach together with a unique double-branch architecture allows the neural network (NN) to learn effectively from historical data and predict inactivation dynamics throughout the whole drying process. In terms of prediction accuracy, our model (R-2 > 0.96) outperforms many other existing models (R2 < 0.6 in most circumstances). By resorting to this model, two optimal SDD conditions have been identified with the resultant terminal solid contents higher than 90 wt% and cell survival ratios higher than 0.65.

Automated summary

A multi-task convolutional self-attention network (CSAN) has been developed for dynamic modeling of probiotics inactivation during single droplet drying (SDD). The model effectively learns from historical data and predicts inactivation dynamics throughout the whole drying process, outperforming many existing models in terms of prediction accuracy. By using this model, two optimal SDD conditions have been identified with high terminal solid contents (>90 wt%) and cell survival ratios (>0.65).