Page 322 - IJB-9-6

International Journal of Bioprinting Rheology-informed machine learning model

and test set split by the ratio of 6:2:2 were performed with where AR is the actual printing resolution and PR is the
an applied random seed and fixed to all the models to avoid predicted printing resolution. This represents the difference
overfitting and estimate learning performance. To assess between the actual and predicted value. Moreover, four
the performance of learning, mean squared error (MSE) different nozzle velocities and five unequal pressure values
was used for loss estimation. were used as the variables of the printing parameters to
Specifically, two classical machine learning algorithms, predict the printing resolution with two conventional
RF and SVM, were prepared based on conventional machine learning models and the RIHML. Specifically,
regression models using ensembles of multiple decision the printing resolution was predicted by different material
trees and hyperplanes with maximum margin, respectively. concentrations of bioinks with a split dataset via CDML
Moreover, the two conventional machine learning models and RIHML. Additionally, a new composition, alginate/
consisted of an input layer, hidden layers, and an output CNC composite bioink, was used to investigate the
layer as shown in Figure 3A and B. Specifically, the possibility of predicting the novel materials not considered
printing parameter-dependent machine learning model in the training set with CDML and RIHML. In the case
uses only printing parameters, such as printing pressure, of CDML, the number of input layer neurons correlate
nozzle velocity, nozzle diameter, and nozzle length. with the number of materials. It means that by adding
Besides, the concentration-dependent machine learning a new material, parameters of the new material, such as
model was incorporated into the input layer consisting of concentration, should be included in the input layer, thus
the printing parameters and the concentrations of each increasing the neurons. To follow the previous structure,
bioink material. the neuron of the CNC existed in the input layer of CDML.
In addition, to provide a visual representation,
Additionally, the RIHML was used with a multi-
input layer model consisting of three input layers with binary images of the printed scaffolds were created using
simulation. To generate the binary image, the printing
hidden layers, concatenated layer, and its hidden layer; strand size obtained from the machine learning model
and an output layer shown in Figure 3C. The rheological was converted into pixels. Using the converted strand size,
properties, such as the viscosity and storage modulus, the single-pixel lines representing the actual printing path
were used as the first and second input layers of the were dilated considering printing directions and angles.
RIHML, respectively; and the last input layer consisted of Afterward, the resulting binary values were converted
the printing parameters. The hidden layers of each input to the binary image to accurately simulate the printed
layer were calculated separately and concatenated with the scaffolds based on the predicted printing strand size.
output layer. The final output was derived with a second
hidden layer whose input is the concatenated output of the 3. Results
first hidden layer.
3.1. Rheological properties of hydrogels
2.6. Prediction of the printing resolution In this study, the viscoelastic properties including viscosity
The machine learning models predicted the printing and storage modulus were assessed and implemented into
resolution with different cases involving new printing a rheological dataset of the bioinks. As shown in Figure 4A,
parameters, different concentrations, and different bioink the viscosity for ten bioinks was prepared to compare the
compositions. More particularly, a dataset was created by viscoelastic properties. Specifically, the shear-thinning
multiple compositions of F127, gelatin/XG, and alginate/ behavior was observed in the flow curves of all bioinks,
CaCl . Splitting was employed to train the model, except and the F127-based bioinks exhibited the highest viscosity
2
for the variable required to perform the prediction in the compared to other bioinks. The flow curves of gelatin-
pre-training stage. Specifically, the independent datasets and alginate-based bioinks are relatively low and partially
in the pre-training stage were provided for each machine overlapped while increasing shear rates. Consequently, the
learning model. They were divided into three parts viscosities for various bioinks were compared with each
(training set, validation set, and testing set) to train the other at the shear rate of 100 1/s as shown in Figure 4B.
model and estimate learning performance. The trained The storage modulus was also measured to investigate the
models have predicted printing resolution with a dataset mechanical strength of the bioinks (Figure 4C and D).
consisting of excluded variables in the training dataset. The Similar to the viscosity assessment, the highest shear moduli
error criterion evaluating printing resolution accuracy was were shown with the F127-based bioink. Additionally,
evaluated using Equation I. viscosities and storage moduli of gelatin- and alginate-based
bioinks were correlated with the increasing concentration
AR PR− of gelatin/XG and CNC, respectively. However, while the
Error%() = ×100 (I)
AR shear moduli of alginate-based bioinks crosslinked with

Volume 9 Issue 6 (2023) 314 https://doi.org/10.36922/ijb.1280