PREDICTION OF CONCRETE MIX COMPRESSIVE STRENGTH USING STATISTICAL LEARNING MODELS

Abstract

Laboratory analysis is expensive. In addition, recent developments in data mining techniques present a great opportunity for researchers to utilize available experimental data in a more productive way by building forecasting models. Today, there are several data-driven learning models that can predict outputs using input variables. Relevance vector machine is emerging statistical learning algorithm used in prediction of complex systems. Random Forest model is a novel algorithm used in a regression where it can assess the importance of input variables in a robust manner for a large number of variables. The two algorithms are developed and compared to concrete mix properties. Adding Silica to concrete mix produce high early strength material, which is a desirable feature. We analyse the results of 90 experimental samples that tested the effect of adding Silica and other physical properties on the compressive strength of concrete. Several silica/cement mixing ratios are evaluated and compressive strength was measured after 3 days and 28 days. Concrete strength increase is proportional with the increase of silica/cement ratio. In this study, the Random Forest model was developed to predict compressive strength using input variables and compared to Relevance Vector Machines model. The R2 for Random Forest model is high at 0.97 but less than Relevance Vector Machines model at 0.989. However, the Random Forest model evaluated variables importance and indicated the curing time and milling time has a higher impact than increasing silicate percent.

Keywords

• Compressive strength
• Concrete
• Milling time
• Prediction model
• Random forest
• Rattle
• RVM
• Silica percent