A Generalized Linear Model for Electrospun Nanofiber Diameter: From Inferential Effects to Predictive Validation

Electrospinning is a versatile technique to produce polymer nanofibers for drug delivery, cellular scaffolds, filtration, and other applications whose performance is strongly influenced by morphology. In previous work, inferential and predictive generalized linear models (GLMs) where used to describe the electrospun fiber diameter ln(Y) from solution and process parameters, including thermodynamic polymer-solvent compatibility descriptors, such as relative energy difference (RED) and the Flory-Huggins interaction parameter (χ), as well as voltage, temperature, feed flow-rate, velocity, tip-to-collector distance, and other electrospinning variables. The inferential analysis identified natural logarithm transformation ln(Y), a Gamma distribution, and an identity link (μ), as the best statistical structure, which was later used to train a prediction model with the same structure, achieving a RMSE = 0.5862 log-nm, Corr² = 0.7803, and MAPE = 7.75%. This model was used to validate electrospu diameter of PVA nanofibers at different voltages, reaching a best-case percentage error of 3.60%.

The present study extends that validation beyond voltage by evaluating targeted variations in RED, χ, solution concentration, feed flow-rate, and tip-to-collector distance. Inferential β coefficients were used to estimate expected percentage changes in Y, while scanning electron microscopy (SEM) micrographs provided experimental diameter measurements. These SEM measurements were compared with both GLM-predicted diameters and β-derived percentage changes.

For the inferential model, the strongest consistency was observed for high-information predictors: RED, χ, and concentration showed discrepancies of 8.53%, 9.34%, and 7.02%, respectively, between β-derived changes and SEM-based diameter variations. For the predictive model, RED and χ showed the lowest percentage errors between predicted and SEM-measured diameters, close to 2.50% in both cases. In contrast, lower-information predictors and reference-dependent transformed terms, such as tip-to-collector distance, showed larger deviations, with predictive errors above 15.00%. These results support that inferential and predictive modeling deviations are dependant on the information provided by each predictor.