Towards an interpretable machine learning model for electrospun polyvinylidene fluoride (PVDF) fiber properties

Abstract

A robust understanding of structure–property relations of electrospun fibers is vital for device design. However, these relationships are inherently complex and hard to model using data from limited trial and error experiments. Machine learning has emerged as an efficient approach to model multidimensional relationships but fundamentally require diverse data to learn these relationships from. In this study, we present a novel Electrospun Fiber Experimental Attributes Dataset (FEAD) by collating experimental data from literature, developing new features, and complementing with our own experiments. Fiber diameter, a key parameter for controlling electrical and thermal properties of electrospun polyvinylidene fluoride (PVDF) polymer, is modeled using a large number of solution and electrospinning process experimental parameters using a multi-model machine learning approach. This is complemented with a model-agnostic interpretable game-theoretic approach to identify the relative and absolute relationships between the variables. Experimental attributes such as feed, polymer concentration, Flory-Huggins Chi parameter, and relative energy difference were found to be most impactful for modeling fiber diameter. This study overcomes several limitations in existing literature such as non-availability of meta datasets, application of latest machine learning techniques, and state-of-the-art approaches for interpreting these “black box” models, thus bridging the gap between experimental and computational studies. This improved ability to generalize structure–property relationships across any PVDF-polymer solvent system presents a promising ability to reduce expensive lab testing required for developing PVDF fibers of desired mechanical and electrical properties.