Mondragon Innovation | Humanity at work – Prediction of Polymer Creep Moduli using Machine Learning
Mondragon Innovation | Humanity at work
Sector: Industry
Business Case
Polymers gradually lose their stiffness over time when subjected to a constant load, a phenomenon known as creep. Tests to measure this phenomenon are very time-consuming and expensive, meaning the industry requires alternative techniques. Once a database was obtained, we explored the usefulness of Machine Learning for creep prediction.
Objectives
Predict creep modulus values over long periods of time with the highest possible accuracy, reducing testing time. In addition, obtain interpretable and generalisable models, and understand which are the parameters that most influence the phenomenon.
Use case
First, download, clean and build the database, which comprises more than 450 polymers and more than 5000 creep curves. Then, fit and validate regression models of increasing complexity, obtaining the best predictions with gradient boosting models (LightGBM).
Infrastructure
On Premise
Technology
Machine learning and deep learning
Data
Dataset in table format, obtained from the free and open-source campusplastics.com database.
Resources
The project is part of a research line with more than 5 researchers. The database download was outsourced, while for the rest we relied on the work of a full-time researcher for 2-3 months and the help of an expert from Mondragon Unibertsitatea (40 hours).
Difficulties and learning
On the one hand, the difficulty of obtaining a complete, faultless and varied (representative) database. On the other hand, the difficulty of being able to trust the predictions, i.e., to estimate the predictive capacity of the models in a realistic manner. In addition, the organisation of the work and the recording of the experiments was fundamental.
KPIs (business impact and metrics of the model)
To assess and compare model accuracy, we employed the coefficient of determination score (R²). For a realistic evaluation of generalisability, we used nested cross-validation (Nested/Double CV), aligning with best practices in materials informatics for data splitting.
Funding
Mondragon funding for the financing of the collaboration with Mondragon University.
Collaborators, Partners
Carlos Cernuda (University of Mondragon)
