SpinCastML─An Open Decision-Making Application for Inverse Design of Electrospinning Manufacturing: A Machine Learning, Optimal Sampling, and Inverse Monte Carlo Approach

Abstract

Electrospinning enables the fabrication of micro- and nanoscale fibrous materials with highly tunable architectures, yet its rational design remains challenging due to the strongly coupled effects of polymer chemistry and concentration, solvent selection, and processing conditions. The process is intrinsically multivariate and nonlinear; small variations in formulation or operating parameters can shift electrohydrodynamic regimes and generate distinct fiber-diameter distributions. Because functional performance often depends on distribution shape and variability rather than on mean diameter alone, electrospinning design must be addressed as a distributional and chemically constrained problem. SpinCastML is a distribution-aware and chemically constrained inverse-design framework implemented as an open standalone executable. The platform is built on a rigorously curated data set of 68,480 individual fiber-diameter measurements extracted from 1,778 data sets spanning 16 polymers. To address high-dimensional heterogeneity and material imbalance, the framework integrates polymer-balanced Sobol and D-optimal sampling with systematic benchmarking of 11 machine-learning algorithms. Rather than predicting a single scalar outcome, SpinCastML models complete fiber-diameter distributions, enabling quantitative assessment of process robustness and probability of meeting user-defined targets. A Cubist model trained under balanced Sobol+D-optimal sampling achieved robust global predictive performance (R2 > 0.92) with stable generalization across chemically diverse systems. The embedded Inverse Monte Carlo engine formulates electrospinning as a probabilistic inverse problem, generating chemically feasible polymer–solvent–process configurations that satisfy compatibility constraints while assigning quantified success probabilities. Experimental distributions are reconstructed with R2 > 0.90, and success rates are predicted within <1% absolute error across independent systems. By integrating distribution-level modeling, chemical feasibility constraints, and probabilistic inverse simulation within a reproducible framework, SpinCastML advances electrospinning from empirical parameter tuning toward quantitatively guided, application-driven design.