Page 12 - IJAMD-1-1

International Journal of AI
for Material and Design ML in 3D bioprinting of cultivated meat

by the increasing number of variables, necessitating (mechanical or pneumatic), layer height, extrusion multiplier,
expert knowledge. ML holds the potential to transform print pressure, and infill density. A total of 345 videos
the bioink development process by creating predictive (115 videos per classification) were used for model training
models for printability based on experimental data. and evaluation, with a train/test ratio of 9:1. The DL model
Beyond the biological considerations, printability stands demonstrated an overall prediction accuracy of 94.3%;
as a fundamental characteristic of bioink utilized in 3D a precision value of 87.2% with a recall value of 96.5% for
bioprinting. 33-35 The development of printable bioinks involves “good print outcome,” a precision value of 97.6% with a recall
multiple steps, such as selecting suitable materials for specific value of 92.2% for “under-extrusion print outcome,” and
applications, formulating bioinks with varying concentrations, a precision value of 98.3% with a recall value of 94.5% for
characterizing the rheological properties of bioinks, and “over-extrusion print outcome.”
conducting printability tests on the formulations. 36-40 Each In another study, response surface methodology was
of these procedures demands highly specialized expertise employed to evaluate the relationship between rheological
in its respective domain, consuming significant time and properties and printability. Thirteen bioink formulations were
resources. This complexity impedes progress in developing used to print filaments and optical microscopy images were used
optimal 3D printing bioinks, highlighting the need for a to assess printability based on filament width and roughness.
42
novel approach to address this challenge.
Random forest (RF) classification models were constructed
ML emerges as a potent solution in this context; using three feature sets – rheological measurements and
ML algorithms can uncover intricate relationships and formulation parameters, rheological measurements alone,
patterns that conventional analysis might overlook by or formulation parameters alone. Predictions for filament
processing and scrutinizing extensive datasets containing width showed that training with formulation parameters
diverse material compositions and their corresponding alone resulted in 5 – 10% lower accuracy compared to the
properties. This capability of ML facilitates the creation other two features. Training with rheological measurements
of predictive models that aid in the selection of ideal alone or both rheological measurements and formulation
material compositions to achieve desired properties parameters yielded similar accuracies, demonstrating the
while considering processing limitations. The choice of potential to predict bioink printability accurately with RF
ML approach for material formulation depends on the classifiers trained on rheological measurements alone without
specific goals and challenges associated with the material specifying formulation parameters.
development process. Material formulation involves
crafting and optimizing materials with the desired Another study has constructed ML algorithms using
properties for 3D printing. Supervised learning is suitable decision tree (DT), RF, and DL to predict the printability
43
when labeled data links material compositions with specific of biomaterials with a train/test ratio of 7:3. A total
material properties or performance metrics. Unsupervised of 210 biomaterial formulations were 3D printed using
learning is appropriate for identifying patterns, clusters, or an extrusion-based printing technique under constant
similarities among materials or their properties without conditions (e.g., nozzle diameter, layer thickness, print
predefined labels. A combination of supervised and speed, and print temperature), and the printability of each
unsupervised techniques can be advantageous for material formulation was categorized based on the shape fidelity of
formulation in 3D printing, involving the initial use of the printed structures. All ML methods successfully learned
unsupervised learning to discover complex relationships and predicted the printability of various bioinks based on
in the data, followed by employing supervised learning to their biomaterial formulations. The RF algorithm achieved
build predictive models for specific material properties. The the highest accuracy (88.1%), precision (90.6%), and
potential role of ML in 3D food bioprinting is discussed in F1-score (870%), indicating superior overall performance
the subsequent sections. among the three algorithms, while DL exhibited the highest
recall percentage (87.3%). In addition, the ML algorithms
2.1. Prediction of printability through ML could generate a printability map of biomaterials to guide
Certain preliminary studies demonstrated the prediction bioink development using a standardized combination of
of print outcomes using ML approaches. A study printing conditions (Figure 2).
applied a robust deep learning (DL) model based on In another study, a hierarchical ML model, governed by
an ad hoc optimized neural network to categorize print rheology data, was implemented to enhance the accuracy
outcomes (good, under-, and over-extrusion). This of predicting the printing resolution of constructs created
41
44
model, complemented by a mathematical model, utilized through extrusion-based bioprinting. The process of
a high-definition webcam to generate a dataset of videos, predicting printing resolution involved several sequential
incorporating various parameters such as printing set-up steps, including the preparation of bioinks with varying

Volume 1 Issue 1 (2024) 6 https://doi.org/10.36922/ijamd.2279