Page 317 - IJB-9-6

International Journal of Bioprinting Rheology-informed machine learning model

principally due to the availability of a wide range of usable and the ideal pressure drop in the system, may not coincide
material viscosity, which allows for freedom in material with reality; hence, the accumulated errors can cause a gap
selection when preparing bioink [10-16] . between the predicted and actual results.

However, the advantages of the generous availability of Recently, multiple studies of printability prediction
bioink candidates including hydrogels, extracellular matrix, using machine learning have been reported to overcome
bioceramic particles, and shear thickening materials create the limitations of physical models. Besides, due to the
difficulties in tuning bioink compositions and finding time-consuming and sequential bioprinting procedure
appropriate printing conditions [17,18] . Thus, to overcome from bioink preparation to 3D deposition, the size of the
these challenges and complexities in the preparation of acquired dataset is relatively small. Therefore, most of these
bioinks and extrusion-based bioprinting, printability has studies employed simple supervised machine learning
recently attracted considerable attention. Although the methods, such as support vector machine, decision
definition of printability in extrusion-based bioprinting tree, lasso regression, and ridge regression, to estimate
is still in discussion, it is obvious that better printability printing resolution with the confined variation of bioink
can improve the printing accuracy and shape fidelity of composition and printing parameters [39-43] . Specifically,
the printed constructs, leading to faster fabrication speed a few studies employed artificial neural networks for
and long-term stable functionality [19-22] . Further, recent printability prediction, but its architecture is hardly
printability studies have quantitatively evaluated printing deep with shallow hidden layers and a limited number
accuracy and shape fidelity using the assessment of various of neurons because of the dataset size [44-46] . Moreover, in
outcomes, such as printing resolution relative to the nozzle most of the neural network-based printability prediction
diameter, distance between filaments, pore size and shape models, superficial parameters such as the concentration
in a grid structure, and height of the stacked layers [23-25] . In of each bioink component were used in the input layer of
addition, these studies have demonstrated that printability the artificial neural network. This may significantly limit
is deeply linked to how to control the rheological properties the expandability of the machine learning model since
of bioinks (viscosity, shear modulus, and gelation) and the neural network should be rebuilt and newly trained
the printing parameters (printing pressure, temperature, for every change in the bioink composition. Therefore, an
nozzle size, nozzle length, and nozzle velocity) [26-32] . improved machine learning model, which is versatile with
However, it is still difficult to find the optimal printing various bioink combinations and penetrates the essence
conditions because the assessment of the printability of bioprinting with profoundly related parameters, is
often relies on trial and error with a repeated change of required.
the bioink composition and printing parameters, which is
time-consuming and cost-ineffective. Furthermore, the rheological properties, such as
viscosity and shear modulus, of bioinks have been
Hence, printability prediction is critical to the accurate verified to be closely related to the printing parameters to
and effective fabrication of tissue-engineered constructs optimize the printability of extrusion-based bioprinting
using the extrusion-based bioprinting technique. In several (Figure 1) [47-53] . For instance, even though the viscosity
existing studies, the physical model-based computation of bioink increases, the flow rate can be maintained if the
was adopted for printability prediction [33-38] . More printing parameters such as pressure and nozzle size are
precisely, the physical model of printability prediction was appropriately increased. Even with the same flow rate, the
derived from hydrodynamic equations combined with printing resolution significantly correlates with the shear
the rheological modeling of generalized Newtonian fluid, modulus and nozzle velocity. Despite the deep correlation
mainly power-law fluid. Using the physical model, the between rheology and printability, there have been no
printing resolution of the output filament was simulated studies that actively applied rheological measurements
with multiple printing parameters and compared with the of various bioinks and multiple printing parameters to
actual printing resolution. Although several studies using machine learning.
the physical model reported interesting results, the model
holds many assumptions and simplifications, limiting its Therefore, in this study, a rheology-informed
application in various bioprinting tasks. For instance, the hierarchical machine learning (RIHML) model was
physical prediction model is highly sensitive to the power developed to predict printability in extrusion-based
law index, which can be obtained by the line fitting of bioprinting. Among previously suggested methods to
the measured viscosity. Thus, small errors in rheological quantify printability, the assessment of printing resolution,
measurement and line fitting may have a significant effect which has been widely applied in numerous printability
on prediction accuracy. Additionally, the assumptions in studies, was mainly adopted. To construct a dataset for
the physical model, such as the incompressibility of bioink training the models, the optical images of printed scaffolds

Volume 9 Issue 6 (2023) 309 https://doi.org/10.36922/ijb.1280