Prediction of bead formation in PVDF fiber across different solvent systems using Interpretable Machine Learning

Bead formation is a typical ramification of electrospun fibers during electrospinning. Presence or absence of beads controls the fiber properties, which dictates its usefulness in diverse applications. However, bead formation is a complex non-linear process influenced by solution properties as well as electrospinning process parameters, mostly explored through trial-and-error experiments. Thus, being able to predict bead formation and identify its causal properties could have tremendous techno-economic value by reducing cost of experimentation. This is challenging as these structure-property relationships between experimental features and bead formation are inherently complex and modelling them requires large datasets from diverse experiments with multiple solvents, which are not commonly available. Here, we developed our own Electrospun Fiber Experimental Attributes Dataset (FEAD) dataset, a curated meta-database of experimental data available in literature, supplemented with our own experiments. Combining it with multiple machine learning models, we showed that while an increase in polymer concentration and applied voltage leads to fewer beads, a higher Flory-Huggins parameter supports increased bead formation. Further, we adopted a game theory-based model agnostic interpretation technique called SHAP (SHapley Additive exPlanations) to identify features contributing towards the occurrence of beads and their relative significance. This novel framework successfully predicted bead formation across various PVDF-polymer solvent systems and demonstrates how community meta datasets, cutting-edge machine learning techniques, and model interpretability methods could be seamlessly integrated to reduce the number of experiments required for developing high quality PVDF fibers. © 2024 The Author(s)