Explainable machine learning for the regulatory environment: A case study in micro-droplet printing

Micro-droplet printing remains an invaluable tool for the additive manufacturing industry due to its versatility, efficiency and cost-effectiveness. The design and deployment of a disposable micro-droplet dispenser seeks to facilitate a quick and easy way to alternate between different printing materials (such as ink, metals, bio-materials, pharmaceuticals, etc.) without the need for thorough cleaning or separate delivery systems. While this works well in principle, in practice achieving consistent performance between dispensers can prove challenging. This paper aims to tackle this issue through the use of Machine Learning (ML) in order to (1) build a model that predicts dispenser performance from a set of physical features, (2) explain the model's predictions by relating features to outcome, and (3) suggest criteria that characterize optimal performance. A micro-droplet dispenser was used in conjunction with a drop-watcher for in-situ monitoring of droplet formation. Using this setup, 300 micro-droplet dispensers were evaluated for printing performance while various physical measurements of the dispensers were taken separately using a high magnification digital microscope. Logistic Regression, Random Forest, Support Vector Machine and Gradient Boosting Machine algorithms were used to train and test classification models based on dispenser performance. AUC scores in the range of 0.75–0.85 were observed for the various algorithms while Partial Dependence Plots were utilized to investigate how the physical attributes of the dispenser impacted performance. Importantly, results were presented and interpreted in the context of providing an insight to transparency for ML in the regulatory environment.