Page 324 - IJB-9-6

International Journal of Bioprinting Rheology-informed machine learning model

Figure 5. 3D-stacked bar graphs with printing parameters for (A) full dataset and (B) test dataset with the bioinks of F127, gelatin/xanthan gum, and
alginate/CaCl . Fitting actual values with prediction values with (C) random forest (RV), (D) support vector model (SVM), (E) parameter-dependent
2
machine learning (PDML) model, (F) concentration-dependent machine learning (CDML) model, and (G) rheology-informed hierarchical machine
learning (RIHML) model. (H) Bar graph of average errors for each model.

as shown in Figure 5E, the PDML predicted every test trend with PDML, resulting in inadequate fitting as
resolution to be 1.68 mm, which does not match the depicted in Figure 5D. These models also exhibited large
actual values ranging from 0 to 4 mm while other artificial errors of approximately 40%, as described in bar graphs
neural network models showed reasonable fitting results in Figure 5H. As a result, the prediction of two classical
(Figure 5F and G). The experimental results from the RF machine learning models and PDML imply that it is not
model had a large standard deviation of errors as shown in appropriate for forecasting the printing resolution of
Figure 5C, indicating that the prediction was significantly various bioink types. Furthermore, the predicted resolution
biased. Furthermore, prediction accuracy using another of the rheology-informed model most accurately matched
classical machine learning model, SVM, showed a similar the actual printing resolution.

Volume 9 Issue 6 (2023) 316 https://doi.org/10.36922/ijb.1280