Page 78 - IJAMD-1-2

International Journal of AI for
Materials and Design
Machine learning for gel fraction prediction

Table 4. Performance of the best machine learning models sample size needed for the other hydrogel model. Common
from each feature group in predicting gel fraction properties for the crosslinking of polymer such as monomer
reactivity ratio and extent of reaction can be included in
MAPE (%) SD (%) R²
Feature Group 1 the dataset for a more robust DNN model where the learnt
knowledge can be transferred easily between models.
Support vector regression 3.13 3.75 0.79
Feature Group 2 4. Conclusion
Random forest regression 9.10 9.92 0.02 This study demonstrates the feasibility of predicting
Feature Group 3 the gel fraction of GelMA-PEDOT:SPSS hydrogels
Random forest regression 6.75 8.56 0.41 using ML models based on bioink formulation and
Deep neural network 6.31 5.78 0.54 crosslinking parameters. SVR emerges as the best-
Abbreviations: MAPE: Mean absolute percentage error; R2: Coefficient performing model, with an MAPE of only 3.13%. This high
of determination; SD: Standard deviation. accuracy minimizes the time and material costs typically
associated with optimizing hydrogel properties to achieve
the required gel fraction. Furthermore, by replacing
crosslinking parameters with absorption coefficient, we
demonstrated the potential for estimating gel fraction
without prior crosslinking information. The DNN model
achieved an MAPE of 6.31% for this scenario, indicating
its utility for in situ gel fraction measurements via a UV
detector. This capability can significantly enhance the
fine-tuning of GelMA-PEDOT:SPSS hydrogel properties
during 3D bioprinting by allowing non-destructive,
real-time measurement of the gel fraction. Overall, this
work contributes to reducing experimental costs and
improving the precision of hydrogel crosslinking, enabling
a more efficient process in hydrogel-related research.
Consequently, this accelerates advancements in the field
of tissue regeneration, providing a robust foundation for
future studies and applications.
Future work should explore the relationship between
gel fraction and various hydrogel properties, such as
rheological behavior, mechanical strength, and cell
viability. Integrating these data with the current models
Figure 7. Graphs of predicted values of gel fraction against actual values will enable users to select optimal parameters tailored
for the testing dataset
to specific applications. Furthermore, the ML models
should be validated in a 3D printer to demonstrate its
Besides, this model assumed that no additional layer is effectiveness in optimizing the crosslinking of hydrogel
added on the sample after the crosslinking process. This during 3D printing. This research lays the groundwork for
model is not suitable in a situation where the transmittivity more efficient and effective design of hydrogels, enabling
of the hydrogel is too low, which necessitates curing after advancements in 3D bioprinting and other critical
every layer. This could be solved by utilizing a recurrent applications in biotechnology.
neural network by having the curing information from the
previous layer as the input for the prediction in the new Acknowledgments
layer. None.

Furthermore, this model is limited to GelMA-
PEDOT:SPSS hydrogel. While this concept can be used to Funding
train an ML model for a different hydrogel, it is preferable This research is supported by the National Research
to have a model that can be generalized for all hydrogels. Foundation for NRF Investigatorship Award No.: NRF-
A possible solution is to use transfer learning to reduce the NRFI07-2021-0007.

Volume 1 Issue 2 (2024) 72 doi: 10.36922/ijamd.3807