Predicting Tire Compound Properties Using Machine Learning
Salmikangas, Anna (2025-10-22)
Predicting Tire Compound Properties Using Machine Learning
(22.10.2025)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe20251107105918
https://urn.fi/URN:NBN:fi-fe20251107105918
Tiivistelmä
Safety, performance, and cost are key requirements when developing new tire compounds. In addition, environmental questions related to tire materials are increasingly rising which introduces a lot of requirements for materials used in tires. Compound development takes a lot of time and money as multiple laboratory tests are needed in order to make sure the compound meets the requirements. The purpose of this thesis is to study if machine learning methods could be used with real life experimental data to predict the properties measured in laboratory. Common compound materials and laboratory testing methods are discussed in the theory part.
The experimental part consists of comparison between six different machine learning methods: linear regression, ridge regression, kernel ridge regression, support vector regression, random forest regression, and gradient boosting regression. Models were created and evaluated using leave-one-out cross-validation (LOOCV) and leave-one-group-out cross-validation (LOGOCV). Evaluation metrics were R2, normalized RMSE, and normalized MAE to overcome the different scales and units of the properties. Kendall’s τ was used as dimensionless metric to evaluate how well the models can maintain the ranking in the data.
The data was combined from 9 laboratory experiments with the total of 86 samples. The aim was to predict eleven properties that are commonly used in the tire industry based on the compound material amounts. Studied properties are maximum torque, minimum torque, difference of maximum and minimum torque, cure time, tensile strength, modulus 300, elongation at break %, hardness, tanδ (0°C), tanδ (60°C), and E’ (30°C).
There was no single all-round best model and there were big differences in models’ prediction performance for different properties. On average, kernel ridge regression showed promising results when using LOOCV. The metrics showed poor performance for LOGOCV. This indicates that the machine learning models can be used to predict properties within modelled material distribution but extrapolating models to new materials possess significant challenges.
The experimental part consists of comparison between six different machine learning methods: linear regression, ridge regression, kernel ridge regression, support vector regression, random forest regression, and gradient boosting regression. Models were created and evaluated using leave-one-out cross-validation (LOOCV) and leave-one-group-out cross-validation (LOGOCV). Evaluation metrics were R2, normalized RMSE, and normalized MAE to overcome the different scales and units of the properties. Kendall’s τ was used as dimensionless metric to evaluate how well the models can maintain the ranking in the data.
The data was combined from 9 laboratory experiments with the total of 86 samples. The aim was to predict eleven properties that are commonly used in the tire industry based on the compound material amounts. Studied properties are maximum torque, minimum torque, difference of maximum and minimum torque, cure time, tensile strength, modulus 300, elongation at break %, hardness, tanδ (0°C), tanδ (60°C), and E’ (30°C).
There was no single all-round best model and there were big differences in models’ prediction performance for different properties. On average, kernel ridge regression showed promising results when using LOOCV. The metrics showed poor performance for LOGOCV. This indicates that the machine learning models can be used to predict properties within modelled material distribution but extrapolating models to new materials possess significant challenges.