Page 71 - IJAMD-2-3

International Journal of AI for
Materials and Design Optimization of membrane shrinkage and stability

and filtration systems. Notably, these membranes In contrast to Euclidean distance, the Mahalanobis metric
4-6
often undergo spontaneous and stimulus-responsive provides scale-invariant and directionally sensitive
shrinkage when exposed to thermal, solvent, or moisture- distance measures, thereby yielding ellipses that more
based triggers. Such shrinkage enables functional accurately reflect the underlying multivariate normal
7,8
transformations, such as self-folding structures, yet its distribution of the data. The adoption of a 68% confidence
anisotropy and instability can cause severe deformation, level corresponds to one standard deviation from the mean
critically limiting its reliability in applications such as in the bivariate normal case, thereby capturing the most
tissue scaffolds. 9,10 statistically representative core of each distribution, while
Recent studies suggest that the shrinkage of electrospun reducing sensitivity to outliers and extreme observations.
membranes originates from a gradient prestrain field within This visualization framework supports inverse mapping,
the cross-sectional area generated during fiber formation, i.e., given a desired target output, its location on the plot can
driven by radial differences in solvent evaporation, be used to identify the nearest confidence ellipse, and thus
prestraining from high-speed collectors, and the rapid infer the most probable input parameter configuration that
fixation of polymer chain orientation. 11,12 Upon exposure can produce the specified response. This approach enables
to external stimuli, such as solvents, the shape memory principled, data-informed decision-making in predictive
effect 13,14 activates the prestrain, causing fiber buckling modeling and design selection.
and resulting in macroscopic, anisotropic membrane According to Figure 1, the experimentally obtained
contraction. Despite these insights, current models remain biaxial shrinkage ratios only cover certain discrete
largely qualitative. Shrinkage control continues to rely on regions. The same region can be covered by different
empirical parameter adjustments, lacking a predictive, combinations of processing parameters. Even when using
quantitative framework capable of capturing the non- identical parameters, the degree of result dispersion varies
linear, multivariable nature of the process and its sensitivity significantly depending on the parameter combination used.
to perturbations. 11 How can we optimally select a combination of processing

Subtle variations in electrospinning parameters, such parameters to achieve the desired biaxial shrinkage ratios
as applied voltage, rotation speed of the collator, distance of the electrospun membrane while minimizing variability?
between electrodes, and solvent concentration, can More broadly, optimizing processing parameters in
significantly affect jet dynamics, fiber solidification, and manufacturing is essential not only for achieving target
internal prestrain distribution, resulting in anisotropic performance but also for ensuring product stability
and often unpredictable shrinkage behavior. Due to and consistency across production. In processes such as
15
the non-linear and multivariable nature of these effects, electrospinning, small variations in parameters can lead to
traditional modeling approaches often lack the accuracy significant fluctuations in fiber morphology, internal stress
and generalizability required for effective control. distribution, and functional outcomes. With the recent
16
Moreover, electrospinning experiments are typically labor- emergence of machine learning in membrane science, there
intensive and sensitive to environmental fluctuations, is increasing potential to address these challenges by enabling
17
hindering the acquisition of high-throughput data. As a data-driven process optimization that simultaneously targets
result, available datasets are often limited in size and scope, performance and stability, as demonstrated in recent studies
further constraining the development of robust process on material discovery, process control, and performance
control strategies. prediction. 18-20 However, many artificial intelligence (AI)-
Based on the experimental data reported in literature, based approaches focus primarily on predicting average
11
Figure 1 illustrates the distribution of system outputs along performance, often overlooking process variability and
21
the rotational and transverse axes, with each data cluster robustness. This limitation is particularly evident under
(applied voltage [kV], rotation speed of collator [rpm], small-sample conditions, where models may capture central
22
distance between electrodes [cm], and solvent concentration trends but fail to reflect sensitivity to parameter fluctuations.
[%]) corresponding to a distinct combination of input As a result, predictions may meet nominal targets while
parameters, denoted by unique “Combo” identifiers. To real-world performance remains unstable. Integrating
enable quantitative interpretation of cluster behavior, each both accuracy and stability into modeling frameworks is
group of data points is enclosed by a Gaussian confidence therefore vital for reliable, data-driven optimization in high-
ellipse delineating the 68% probability region. These variability manufacturing scenarios.
ellipses are computed based on the Mahalanobis distance, More broadly, optimizing processing parameters in
which incorporates the full covariance structure of each manufacturing is crucial not only to meet performance
cluster to account for correlations and anisotropic variance. standards but also to ensure consistent product stability.

Volume 2 Issue 3 (2025) 65 doi: 10.36922/IJAMD025260022