Predicting the Recipe: Leveraging Prameterless Self-Organzizing Maps (PLSOM) with Geometrical Polynomial Fitting (GPF) to predict biopolymer composites characterization with small datasets

Problem statement
Biopolymer composites are an interesting class of materials to be explored using robotic 3d printing for architectural applications. They provide the opportunity to incorporate biodegradable waste-stream materials into the built environment. Mixtures and recipes can be tuned across the print to satisfy specific performance requirements within the global design.

However, when considering gradual and continuous grading responding to multiple objectives, rather than discrete compositions, two limitations arise:
1. High dimensional recipe-performance space: Multi-ingredient variation to multi-performance response mapping becomes exponentially complicated.
2. Large quantity of experimental samples: Physical in-lab characterization of continuous recipe permutation becomes exponentially unfeasible due to material and time costs.

Our approach
The project develops an experimental methodology to predict, with sufficient accuracy, the physical performance of all ingredient permutations within a recipe space, using a small physical dataset of lab-samples. By leveraging the associative positioning of Self-Organizing Maps we are able to geometrically fit a Polynomial model which outperforms state of the art predictive models. The low-dimensional mapping also allows us to develop an intuitive interface to navigate the ingredient-performance response and plugs directly within a computational design workflow.