Page 328 - IJB-9-6

International Journal of Bioprinting Rheology-informed machine learning model

cases, including new printing parameters with trained and highly time-consuming [55-57] . Therefore, it is crucial
bioink materials, new concentrations of the trained bioink to develop an efficient machine learning model that
constituents, and untrained bioink compositions with the is suitable for small dataset sizes while ensuring high
new material. For different nozzle velocities, the results prediction accuracy. With the hierarchical architecture of
of the study showed that the RIHML model exhibited the developed model, RIHML can effectively predict the
the lowest error percentage (around 18%) in predicting printing resolution of extrusion-based bioprinting using
the printing resolution. Additionally, the RIHML model small datasets. In this study, the dataset of 537 numbers
showed low error (around 10%) in predicting the printing of bioink rheological properties and printing process was
resolution for different pressures, which is 2-fold and used for training, validation, and testing of the machine
5-fold lower than CDML and PDML, respectively. learning model. Several bioprinting studies employed
Moreover, the printing resolution for different bioink small datasets to optimize printing resolution and
concentrations was predicted, and it was demonstrated parameters using conventional machine learning skills,
that the RIHML model exhibited lower error percentages but their practical applications were limited due to low
than the CDML model for all the different concentrations prediction accuracy, poor expandability, and low training
of bioink constituents such as F127, gelatin, xanthan gum, efficiency [41,42,44,45] . However, the RIHML model can easily
and CaCl . Additionally, the machine learning models generalize and embrace new data, even with a small dataset
2
were used to predict the printing resolution with a new size owing to its intrinsic features in the dataset that are not
material (CNC) added to the alginate-based bioink, which biased to specific bioink, but rather are general. Moreover,
is the most challenging among the three cases. The results due to the potential of data accumulation, if various
of the study showed that RIHML can predict the printing rheological and printing data are additionally collected
resolution with reasonably low errors while the printing in sufficient size for deep learning, prediction using a
resolution was hardly predictable using CDML with rheology-informed neural network with deeper hidden
considerably large errors. layers may be attempted.

Overall, the experimental results indicate that the Although the RIHML model has the potential for
rheology-informed hierarchical model can be a useful accurate and robust prediction of printability, there is still
tool to predict the printing resolution of extrusion-based room for improvement. Due to the generalizability of the
bioprinting. Furthermore, while other studies related bioink properties, a wider range of rheological properties
to the prediction of printability in bioprinting could of bioinks can enhance the prediction accuracy of the
anticipate the printability changes only with limited RIHML model. For instance, in the results presented in
parameters, such as bioink material properties or printing Figure 7C, relatively high errors were observed in F127
conditions, the RIHML model is versatile to predict the with a concentration of 45%. Specifically, this may occur
printing resolution in different conditions of varying because its viscosity and storage modulus were the highest
printing parameters, varying material concentrations, around the upper bound of the rheological data range.
and new bioink compositions [40,43,54] . Additionally, In terms of future work, it would be beneficial to further
the neural network structure of RIHML is based on validate the performance of the RIHML model from other
rheological properties, which can be widely obtained types of bioprinting methods, such as inkjet-based or
from most biomaterials, and it can be trained without laser-assisted bioprinting, to demonstrate the feasibility of
significant alterations of the structure. Therefore, the rheology-based prediction of printability across different
RIHML model is adaptable and expandable compared to bioprinting methods. Additionally, future studies could
the conventional models, and the printing and rheological investigate the potential of the RIHML model in predicting
datasets may be accumulated to enhance the prediction other aspects of printability, such as the extrudability,
accuracy. pore size, pore shape, and shape fidelity of the stacked
layers.
Since the formulation of bioinks and the process of
bioprinting are more complicated and correlated, the 5. Conclusion
prediction of printability in 3D bioprinting has become
more challenging. Recently, there have been attempts In conclusion, this study suggests that the rheology-
related to the prediction of bioprinting printability using informed hierarchical model can be a useful tool for
machine learning. However, unlike other fields such as predicting the printing resolution of constructs fabricated
medical imaging and genetics, 3D bioprinting suffers by extrusion-based bioprinting. Interestingly, the RIHML
from data size, which may hardly be large because the model demonstrated the lowest errors (around 18%) in
preparation of bioinks with various compositions and predicting the printing resolution for different printing
their 3D printing with multiple parameters are sequential parameters such as nozzle velocities and pressures,

Volume 9 Issue 6 (2023) 320 https://doi.org/10.36922/ijb.1280