Page 74 - IJAMD-2-3

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

target
base value f (mean model output), the prediction can be and TD, denoted as R target and R TD . In addition,
0
RD
decomposed as Equation II. tolerance limits ∆ and ∆ are defined to account for
RD
TD
fx() f M i (II) permissible deviations from the target values. These
targets are determined based on the desired
0
i1
dimensional accuracy of the electrospun mat after
This additive decomposition provides a transparent post-processing, ensuring the end product meets
interpretation of how each electrospinning parameter application-specific requirements (e.g., biomedical
affects the model output, enabling a parameter-sensitive scaffolds or filtration membranes). This initialization
optimization strategy. step provides clear quantitative objectives for
2.3. Optimization of shrinkage stability subsequent parameter screening.
(ii) Step 2: Monte Carlo sampling
To identify optimal processing parameters that achieve A large set of candidate processing parameter
target shrinkage ratios with maximal stability, a Monte Carlo combinations is generated via Monte Carlo sampling
simulation 34-38 framework was established by integrating within predefined practical ranges. The considered
shrinkage prediction and shrinkage stability models. The parameters include applied voltage (kV), solution
Monte Carlo method is a stochastic simulation technique concentration (wt%), collector distance (cm), and
that estimates numerical results by performing repeated rotation speed (rpm). Random sampling ensures
random sampling over the parameter space. In this study, comprehensive coverage of the parameter space,
each process parameter, including applied voltage, TPU enabling the identification of non-intuitive optimal
concentration, collector speed, and electrode distance, is combinations that conventional trial-and-error
treated as a random variable within experimentally feasible methods may miss.
bounds. A large number N of parameter sets is generated (iii) Step 3: Shrinkage prediction
from the joint distribution of process parameters to Each sampled parameter set is evaluated using a
explore the multidimensional process space in an unbiased previously trained shrinkage prediction model,
and comprehensive manner. Compared with Bayesian which was constructed based on experimental
optimization or other sequential search algorithms, this data and machine learning algorithms. The model
one-step large-scale sampling can simultaneously evaluate outputs predicted shrinkage ratios in ( ) and
ˆ
RD R
shrinkage accuracy and stability without iterative model ˆ RD
( ) ,
updates, making it particularly suitable for our non-linear, TD R TD capturing the non-linear dependencies
high-variability system with a pre-trained predictive model. between processing parameters and shrinkage
behavior. This predictive approach greatly reduces the
Mathematically, for a target shrinkage function f(x) and
stability metric g(x), the expected values are approximated number of costly experimental trials.
as Equation III, (iv) Step 4: Feasibility filtering
Only those parameter sets whose predicted shrinkage
1 N 1 N ratios satisfy Equation IV,
Eg x()]
Ef x[( )] fx(),[ gx( ) (III)
N k1 k N k1 k | R ˆ − R target |≤ ∆ , R − R target |≤ ∆
ˆ
|
RD RD RD TD TD TD (IV)
where x is the k random sample of process parameters. are retained for further analysis. This filtering step
th
k
Through sufficient sampling, the distribution of predicted eliminates parameter combinations that would
shrinkage ratios (%RD, %TD) and stability coefficient produce excessive dimensional deviation, thus
(CIW) can be estimated. narrowing the candidate pool to only potentially
By filtering all samples that meet the target shrinkage viable solutions.
tolerance, the Monte Carlo method provides a numerical (v) Step 5: Stability assessment
approximation to the feasible design domain. The For each feasible parameter set, shrinkage stability is
subsequent ranking by predicted CIW corresponds to an evaluated using a trained stability prediction model.
optimization over the estimated probability distribution The stability metric is expressed as the predicted CIW,
of outcomes. This enables the selection of parameter which reflects the sensitivity of shrinkage behavior
combinations that maximize process robustness while to process fluctuations. A smaller CIW corresponds
satisfying the desired biaxial shrinkage performance. The to a more robust process configuration, less prone to
method proceeds as follows: variation due to environmental changes or equipment
(i) Step 1: Initialization drift. This step ensures that the selected parameters
The optimization process begins by specifying the are not only accurate but also reproducible in real
target shrinkage ratios in the radial direction (RD) production.

Volume 2 Issue 3 (2025) 68 doi: 10.36922/IJAMD025260022