Machine learning for rubber development
For most of its history, compound development in the rubber industry has been a craft. The chemist sketches a recipe based on experience and an understanding for the chemistry, the batch is mixed and cured, properties are measured and adjustments are made. Round after round of iterative testing follows before the compound finally meets specification.
Machine learning (ML) offers a different path. Models trained on historical compound data can estimate mechanical properties directly from the formulation, without anything ever entering the mixer. The idea is not new; it has been around in the materials informatics literature for over a decade (ref. 1), and recent work has applied it specifically to natural rubber tensile prediction (refs. 2 and 3). What has remained stubbornly difficult, though, is industrial adoption. The reasons are familiar enough to anyone working in a compound development laboratory: The data demands are large, the models can feel opaque, and there is rarely a clear warning when a prediction should not be trusted.