Viability and prediction of encapsulated Trichoderma spp. isolates using machine learning: a multi-layer...

The development of microbial bioproducts requires understanding the stability and viability of cells during storage. Accordingly, this study evaluated the viability of capsules containing different strains of Trichoderma spp. isolated from native plants of the Piauí Cerrado, using an integrated approach combining microencapsulation and computational modeling. Capsules were produced by ionic gelation in a sodium alginate matrix, and fungal viability was monitored by plating and conidial counting over 60 days of storage. The resulting data were used to train a Multilayer Perceptron (MLP) Artificial Neural Network, structured with an ELU activation function, L2 regularization, and dropout to prevent overfitting. The model was also able to generate projections beyond the experimental period; however, these estimates are strictly exploratory in nature and serve only as a proof of concept for the application of MLPs to describe potential trends beyond the 60 days observed. Within the experimental window, strains UFPI07, UFPI10, UFPI11, and UFPI16 maintained populations above 7 log CFU mL^- 1, indicating greater relative stability, whereas UFPI06 and UFPI18 exhibited a more pronounced decline. The combination of encapsulation and machine learning demonstrates potential as a tool to support preliminary estimates of temporal behavior, optimize formulations, and reduce experimental costs, representing an exploratory methodological advance for predictive microbiology and bioproduct development.