Page 73 - IJAMD-1-2

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

capturing non-monotonic relationships between variables, set (200 samples), validation set (58 samples), and testing
which are prevalent in systems when the variables interact set (29 samples). The training set was used for training for
non-additively. The hierarchical structure of decision trees the ML model, and the performance was tested against the
is particularly useful for decomposing the decision process validation set. The testing set was kept untouched until
53
into a series of simple rules, offering invaluable insights into the end of all training to verify the effectiveness of the
the relationships between parameters. The mean squared models. Both MAPE and coefficient of determination (R )
2
error (MSE) criterion was used for splitting, employing the were used as the performance criteria for model validation.
“best split” approach to optimize the decision tree. MAPE was selected over other criteria such as MSE or
mean absolute error as the percentage error gives a better
2.4.4. RFR
interpretability on how much the prediction deviates from
RFR constructs multiple DTR using subsets of the dataset the real value. Meanwhile, R is dimensionless and can
2
and averages their outputs to predict the target variable. This be used to compare the performance of different models,
52
approach improves accuracy and reduces overfitting compared as an R closer to 1 for a model indicates a better fit. All
2
to a single DTR. However, RFR models are challenging to the hyperparameters stated in sections 2.4.1 to 2.4.5 were
visualize and interpret. The criterion for splitting used is MSE. tuned with grid search method, while the default value was
The number of trees in the forest was set as 15. used for the unmentioned parameters.
2.4.5. DNN Each of the ML techniques described in section 2.4 was
DNN outperforms traditional regression and decision tree trained against the gel fraction with feature Groups 1, 2,
methods in conventional ML technique by automatically and 3 as the input. The best ML technique for each feature
extracting complex features, handling non-linear relationships, group was verified with the testing set.
and scaling effectively with large and high-dimensional 3. Results and discussion
datasets. While regression models such as LR or SVR are
limited by their linear assumptions and decision trees-based Figure 4A demonstrates that all variables in the dataset
models such as DTR or RFR often risk overfitting, DNN influence the resulting gel fraction. Higher concentrations
leverages multiple layers of non-linear transformations of GelMA and LAP facilitate faster curing of the ink
to capture intricate patterns in data. However, DNN will (P < 0.05 for both in one tailed t-test), thereby increasing
perform worse than conventional ML techniques when the the gel fraction. In addition, greater UV power and longer
sample count is low. exposure duration enhance the UV energy received by the
The architecture of the DNN model used in this work photoinitiator (P < 0.05 for UV power, P < 0.15 for UV
is illustrated in Figure 3B. It has four hidden layers, with duration), leading to more crosslinking. Conversely, the
16 nodes in the first hidden layer, 32 nodes in the second higher concentration of PEDOT:SPSS in the ink obstructs
hidden layer, and 16 nodes in both the third and fourth the UV light from activating the photoinitiator, resulting
hidden layers. ReLU activation was used for the first, second, in reduced crosslinking and a lower gel fraction (P < 0.25).
and third hidden layers. The fourth hidden layer utilized a The Spearman correlation shown in Figure 4B
linear activation instead. To prevent overfitting, dropout corroborates these observations. According to the
with a rate of 0.2 was applied between each hidden layer. correlation data, PEDOT:SPSS concentration has the most
The model was trained using standard backpropagation significant impact on the gel fraction, followed by the
and optimized with the Adam optimizer. The loss function concentrations of GelMA, UV duration, UV power, and
used was mean absolute percentage error (MAPE). LAP. The correlation coefficients are not particularly high,
2.5. ML training procedure with the rank for PEDOT:SPSS concentration being −0.47.
This moderate level of correlation justifies the application
The distribution of the gel fraction for 287 samples obtained of ML to better predict the gel fraction.
from the experiment in this work is depicted in Figure 3C.
There are six features and one label in the dataset. The six It is also noteworthy that the absorption coefficient of
features are GelMA concentration, LAP concentration, the samples is moderately related to the gel fraction, with a
PEDOT:SPSS concentration, UV power intensity, UV correlation coefficient of −0.42. As observed in Figure 4C,
duration, and absorption coefficient, with the gel fraction the UV intensity measured increases over time while the
as the label or output for the ML model. Every feature in absorption coefficient decreases as the sample cures. This
the dataset was normalized with a MinMaxScaler, which observation suggests the potential for using the absorption
transformed every feature to a range of 0 – 1. The sample coefficient to perform in situ predictions of the gel fraction,
was then randomly split in a 70:20:10 ratio into training a concept which will be further elaborated in section 3.2.

Volume 1 Issue 2 (2024) 67 doi: 10.36922/ijamd.3807