Abstract text (incl. figure legends and references)
Drop-on-demand technology provides a cost-efficient approach for high-throughput bioprinting with high cell-viability rates. However, these processes face multimodal challenges in terms of process accuracy and stability.
This study aims to develop a system for real-time data acquisition of droplet parameters, for various bioinks and to provide immediate feedback to a higher-level control system to improve process accuracy and stability.
In order to map the entire droplet deposition process, several sensors (pressure, dynamic vision sensor, camera) have been integrated into an existing bioprinting system. Conventional (decision trees), advanced stochastic approaches (bayesian optimization), and deep learning models (reinforcement learning) were compared for data merging and evaluation logic.
The embedded sensors provide a full representation of the printing process in real time. The use of bayesian optimization and clustering of the processed data coupled to it proves to be stable and is superior to the other approaches (decision trees and reinforcement learning) in terms of flexibility, accuracy and speed. The system provides real-time feedback (< 40 ms) on droplet shape, size, and speed, allowing for the application in a process control system. Tests using hydrogels (e.g. agarose) with varying concentrations, printing parameters (0.1 – 1 bar) and valve opening times (300 µs – 1000 µs) revealed high accuracy with respect to the detection of droplet velocity and eccentricity. However, the volume calculation shows a high average relative standard deviation of ~6% due to the two-dimensional approach.
The presented work shows a good suitability of the applied approach for monitoring and control of the DOD process. In the future, the tool could be used for predictive maintenance applications, e.g. identification of nozzle clogging, for real-time optimization of printing conditions, or to characterize material properties (e.g. viscosity) on the fly.