A recent master’s thesis by Anna Salmikangas from the University of Turku explores whether machine learning can support the lengthy and complex process of tire compound development.
Why Explore Machine Learning?
Anyone working in tire material development knows how much trial and error is involved in the work. Safety, performance, and cost are always crucial, and environmental demands have added even more complexity to the equation. As a result, laboratories invest significant time and resources into small but essential experiments to understand how different ingredients interact.
Because there is no universal formula that predicts how tire compounds behave, laboratory testing will always remain necessary. However, if certain properties could be estimated in advance, engineers could focus their efforts on the most meaningful experiments and reduce unnecessary testing. This is why intelligent, data-driven tools are becoming increasingly appealing in the tire industry.
What the Thesis Aimed to Do
The thesis aimed to examine whether machine learning methods, when applied to real experimental data, can help predict the material properties typically measured in the lab. It also reviewed the common materials used in tire compounds and the standard testing methods applied in the field.
Salmikangas worked closely with the Tire & Material Development team at Black Donuts’ new InTire Labs to analyze existing datasets, develop predictive models, and create a practical tool capable of estimating compound properties from historical test data.
How the Research Was Done
The dataset used in the study was derived from nine different laboratory experiments, resulting in a total of 86 samples containing 81 distinct materials, each with recorded PHR (parts per hundred rubber) values. The aim was to predict eleven key properties commonly used in tire development: maximum torque, minimum torque, the difference between them, cure time, tensile strength, modulus 300, elongation at break, hardness, tanδ at 0°C, tanδ at 60°C, and E’ at 30°C.
To test predictability, six machine learning approaches were compared, from simple linear regression to advanced methods such as support vector regression, random forests, and gradient boosting.
Each model was evaluated using two strategies: leave-one-out cross-validation (LOOCV) and leave-one-group-out cross-validation (LOGOCV). Because the studied properties vary in scale and units, the results were evaluated using R², normalized RMSE, and normalized MAE. In addition, the thesis used Kendall’s τ to assess how well the models preserved the ranking of samples, which is crucial when comparing the performance of different formulations. Because the study compared properties on different scales, the use of dimensionless metrics proved helpful, making it easier to compare model performance across properties.
LOOCV evaluates a model by leaving out one sample at a time.
LOGOCV tests how well the model performs on entirely unseen material groups.
What the Study Found
The results were encouraging. Machine learning demonstrated the ability to estimate several properties with useful accuracy. It performed best when materials were familiar, a common expectation in any field where models learn from historical trends. Some properties were easier to predict than others, revealing which areas of the dataset were strong and where the data was thin or uneven.
One of the most valuable insights came from the limitations of the model itself. Differences between LOOCV (strong performance) and LOGOCV (weaker performance) showed that today’s models work well within familiar material spaces but struggle when extrapolating to new materials. This highlights exactly where more experimental data is needed.
What This Means for the Future
The research suggests that machine learning can make tire material development more efficient by helping engineers use historical data more intelligently. It supports faster development cycles, more informed decision-making, and more targeted laboratory testing, especially important as the industry moves toward sustainable materials and entirely new formulations.
With machine learning, engineers can better understand patterns in existing data and gain early insights into how a new formulation might behave before conducting physical tests. As more high-quality data becomes available, prediction accuracy will continue to improve and error margins will shrink.
The new InTire Labs facility will play a key role in this development by generating cleaner, more consistent data from materials and tires, making intelligent methods even more effective in future research.
