Uncertainty-aware neural networks for autonomous real-time correction in bioprinting

The inherent instability of biomaterials in bioprinting often leads to low geometric fidelity and frequent print failures. While conventional strategies rely on passive pre-process optimization, they remain incapable of compensating for stochastic perturbations in real-time. In this work, an intelligent closed-loop control system is introduced to enable autonomous correction in extrusion-based hydrogel bioprinting. At its core is a multi-task deep neural network, the Printing Reliability In-Situ Monitoring Network (PRISMNet), which simultaneously predicts a comprehensive set of corrective control parameters—including both continuous and categorical outputs—directly from live video streams. Critically, the model integrates uncertainty quantification, enabling decisive correction when confident while maintaining a conservative stance in the face of ambiguity to prevent model-induced errors. By integrating this model with a Kalman filter-based correction module, we established a robust feedback loop that dynamically adjusts printing parameters. Experimental results demonstrate that this active control strategy significantly outperforms open-loop systems, effectively rectifying common defects such as material accumulation and discontinuity. The system successfully improved the geometric fidelity of complex hydrogel constructs, validating its practical utility and providing a scalable solution for autonomous bioprinting. • An uncertainty-aware neural network for autonomous bioprinting correction. • NN combined with a Kalman filter ensures stable and robust closed-loop control. • Rapidly corrects defects, significantly improving fidelity of complex structures. • The system generalizes to novel bioinks (Pluronic F127) unseen in training.