Page 316 - IJB-8-4

Deep learning for EBB control
aspects of the EBB process, including material, scaffold strategy was shown to be the best fit for anomaly detection
geometry, and printing apparatus that determines a priori on the provided dataset .
[17]
if the printing process will be successful or not) [7-9] . Despite the good results of these few preliminary
However, there is often a delicate balance between these studies, much research is still needed for the
different requirements, which results in careful and implementation of a robust, AI-based quality control
time-consuming optimization of the material properties, system of the EBB process. First, this system should
printing parameters, and scaffold geometry to obtain good be able to analyze the printing outcome and change the
print quality results. Furthermore, since no standardized printing parameters without any manual involvement to
method is currently available for the printing parameters have a completely autonomous control loop. Furthermore,
optimization, the results from this trial-and-error process the quality assessment should focus on the full, 3D
may not be reproducible across labs. Altogether, these scaffold shape, and not only on a few, initial layers,
limitations pose a significant obstacle to the translation which are not representative of the final shape. Finally,
of a promising ink/bioprinted product to more impactful the system should also be able not only to optimize the
clinical applications, as it is more difficult to comply to printing parameters but also monitor the printing session
relevant healthcare-related standards . for quality assessment and potential on-the-fly correction
[10]
In the last few years, artificial intelligence (AI) of working parameters.
technologies have been proposed to improve the control Considering these challenges, we present a novel
and automatization of the complex EBB process [11-14] . In approach for parameter optimization and in-process
this regard, Fu et al. optimized the printing settings in monitoring of the EBB process using a combination of:
EBB using the support vector machine (SVM) machine (i) a robust DL model (based on an ad hoc optimized neural
learning (ML) model. In their research, the authors network) which can classify the print outcome into one of
established an initial set of printing settings to be adjusted, three classes (i.e., good, under, and over extrusion) and (ii)
including cartridge temperature, layer height, and needle a mathematical model (already published in Bonatti et al. )
[7]
gauge for a Pluronic F127-based biomaterial ink. The to automatically tune the printing parameters toward the
authors printed three-layer grid-like designs using various optimal combination. Briefly, we built a comprehensive
combinations of these parameters, and they assessed the dataset by taking videos of the EBB process from a frontal
faithfulness by measuring the line width and comparing it view while printing multilayer scaffold shapes. To model
with the theoretical one. A 3D map was generated using multiple printing outcomes, for each print, a different
the SVM model, indicating the parameter combination combination of relevant parameters (e.g., layer height,
that could yield higher quality prints within a given flow, infill, color of the material, and printing set-up) was
probability threshold . In another recent work, Rubero tested. The dataset is publicly hosted in Zenodo (https://
[15]
et al. proposed an iterative method based on Bayesian doi.org/10.5281/zenodo.7024007). After frame extraction
optimization for the optimization of EBB parameters and pre-processing, the dataset was used to train and
(e.g., biomaterial ink composition, reservoir and bed comprehensively evaluate a CNN architecture with high
temperatures, extrusion pressure, and printing speed). classification accuracy. Finally, we demonstrated how the
Briefly, a set of lattice structures with random parameters classification model could be used in a feedback loop to
were printed to begin the optimization process. These monitor the printing outcome and automatically optimize the
first prints were visually evaluated to compute a quality printing parameters through a series of consecutive prints.
index and employed to create a probabilistic model based The paper is organized as follows: given the
on a Gaussian process. A new set of printing parameters limited amount of literature on AI-based quality control
was proposed by the model and the experimenter used for EBB, in Section 2, we give the reader both a brief
these parameters to create a fresh batch of prints and introduction on ML and DL, as well as an analysis on
corresponding scores. The process was repeated until the current literature on the use of ML for the broader
convergence . field of additive manufacturing. Through this section,
[16]
The use of more sophisticated ML models for the reader will obtain a more comprehensive overview
anomaly identification and localization was studied by of recent techniques, algorithms, and applications, which
Jin et al. Specifically, the authors printed several infill serves as a basis for the work described in the remaining
patterns, such as grid, rectilinear, gyroid, and honeycomb, sections. In Section 3, we describe the main methods used
at various printing speeds while taking top-down pictures to build the dataset and train and evaluate the DL model,
of the first layer. The photos were divided into several as well as how this can be used in a feedback loop for
categories, such as good prints and prints with errors. online monitoring and automatic parameter optimization.
To automatically classify the prints, various ML models, In Section 4, we detail and discuss the results achieved
including SVM and a deep learning (DL) convolutional during the experiments, while Section 5 will be focused
neural network (CNN), were tested. The CNN-based on the conclusions and future developments of the work.

308 International Journal of Bioprinting (2022)–Volume 8, Issue 4