Abstract/Summary

Machine-Learning models thrive on data. The more data available, or creatable, the more defined is the problem representation, and the more accurate is the obtained prediction. This presents a challenge for physical, material datasets, specifically those related to fabrication systems, in which data is tied to physical artefacts which necessitate fabrication, digitisation and formatting to be used as input for predictive models. In this paper we present a design-based methodology to producing a material dataset for statistically modelling the curing of cellulose-based 3d-printed components, as well as associated methods for geometric data encoding, tolerance-informed data augmentation and statistical modelling. The focus of the paper is on the digital workflows and considerations for dataset composition - the material case of 3d-printing cellulose is secondary. We use a built 3d-printed demonstrator wall as a material dataset, through which we generate datapoints that stem from a real design-scenario and inform the fabrication model. Using a feature-engineering approach, select geometrical features are encoded numerically. We perform statistical analysis on the data, and test different shallow models and neural networks. We report on the successful training of a Polynomial Kernel Ridge Regressor to predict the vertical shrinkage of the pieces from wet print to dry element.