Page 327 - IJB-9-6

International Journal of Bioprinting Rheology-informed machine learning model

Figure 8. Prediction of printing resolution using the trained models with different concentrations of CNC incorporated with 2% alginate. (A) The number
of results for alginate/CNC composition at different velocities. Fitting actual values with prediction values with (B) CDML and (C) RIHML. (D) Bar graph
of average errors for each model. (E) Error values for different bioinks formulations and prediction models. Visual comparison between (F) actual image
of printed alginate/CNC composition and (G) simulated image using the printing resolution predicted by RIHML. Abbreviations: CDML, concentration-
dependent machine learning; CNC, cellulose nanocrystal; RIHML, rheology-informed hierarchical machine learning.

RIHML existed within the range of the axis as shown in resolution of the alginate/CNC composition predicted by
Figure 8C. To quantitatively elucidate this phenomenon, RIHML are presented in Figure 8G and agreed well with
Figure 8D and E can differentiate the average prediction the actual images.
accuracy between the concentration-dependent model
and the rheology-informed model. The RIHML method 4. Discussion
can adequately predict the printing resolution of bioink
with new material, but CDML predictions are unreliable This study reports the application of a rheology-informed
and have significant errors. Specifically, the concentration- hierarchical model to enhance the prediction accuracy
dependent model shows approximately 10-fold errors of the printing resolution of constructs fabricated by
compared to the rheology-informed model using the same extrusion-based bioprinting. Specifically, five different
prediction dataset. This result implies the performance of machine learning models, including the RIHML model
RIHML is less affected by the bioink composition, even as well as two classical machine learning models (RF and
with new materials. Furthermore, to visually compare the SVM) and the conventional models based on artificial
strand size of actual printing and prediction using RIHML, neural networks (concentration-dependent model and
the binary images of the printed scaffolds were created printing parameter-dependent model), were trained
using simulation and compared with their actual images. and tested using a small dataset of bioink properties and
Figure 8F shows actual images of the printed alginate/ printing parameters. More precisely, the models were
CNC scaffolds. The simulated images using the printing used to predict the printing resolution in three different

Volume 9 Issue 6 (2023) 319 https://doi.org/10.36922/ijb.1280