Material informatics and impact of multicollinearity on regression model for fatigue strength of steel

In the last few decades, the advancements made in material characterisation equipment and physics-based multiscale material modeling have generated vast database in the field of Material Science and Engineering. This has inspired material innovators to attempt predicting mechanical properties of synthesised materials using big-data so as to reduce the cost, time and effort for materials innovation. However, the impact of collinerarity has always been a matter of concern in emperical research, specially in such predictions of mechanical properties. In the present work, we revisit NIMS database for steel and study the effect of multicollinearity on regression based models for predicting fatigue strength for the material. We use an iterative scheme to isolate highly correlated parameters contributing in determination of the fatigue strength of the steel. We then construct a regression model using only the non-correlated parameters to make the model more efficient computationally. Our results show that the regression model built after consideration of multicollinearity of the variables provide better performance in comparison with regression model built without consideration of the same.