Machine learning predicts nanofibre thickness from electrospinning settings, tested against new scaffolds

The work

Technical · from the record

The authors compiled 68,538 fibre-diameter measurements from 1,778 electrospinning studies covering 16 polymers and, for each polymer separately, fitted seven machine-learning regressors that map six routinely reported process parameters to fibre diameter. Aggregated cross-validated predictions, plus a residual-bootstrap step, give a predicted diameter distribution rather than a single mean, and the models are served through a Shiny web application together with variable-importance, SHAP and correlation diagnostics. Across polymers the non-linear and local learners reached higher coefficients of determination than the linear baselines, and solution concentration ranked as the most influential predictor. In a validation case with newly electrospun PVA scaffolds, predicted and measured distributions were not distinguished by the statistical tests for the Spraybase system (Kolmogorov–Smirnov p = 0.13, overlap coefficient 84.11%), while for the TL-01 system the same tests indicated significant differences (Kolmogorov–Smirnov p = 1.4 × 10⁻⁵, overlap coefficient 73.97%).

How AI was used

Fibre-diameter observations and processing conditions were extracted from the literature and from two published datasets, parsed numerically, filtered for missing or non-finite values, and split into polymer-specific subsets in which zero-variance predictors were removed and numeric predictors z-normalised within each resampling split. For every polymer, seven supervised regressors — ordinary least squares, elastic net, decision tree, random forest, radial-basis SVM, k-nearest neighbours and MARS — were trained on six process variables (solution concentration, needle diameter, rotation speed, voltage, flow rate, tip-to-collector distance) with the fibre diameter as the target. Hyperparameters were tuned on fixed grids in an inner five-fold cross-validation with two repeats, while an outer leave-one-study-out scheme produced out-of-fold predictions and fold-wise R, RMSE and MAE. The aggregated out-of-fold predictions supply the displayed prediction distribution; for a user's settings the fitted model returns a point prediction, which is combined with resampled cross-validated residuals in a 100-realisation Monte Carlo residual bootstrap to form a conditional predictive distribution. Model behaviour was examined with caret's native variable importance and with fastshap Shapley attributions computed by perturbing features on background rows from the polymer-specific training set. The predicted distribution was then compared with diameters measured from newly electrospun PVA scaffolds imaged on two scanning electron microscopes.