Interpreting the Strength Activity Index of Fly Ash with Machine Learning

Fly ash from the coal combustion at electric plants is commonly used for partially replacing portland cement in concrete production. Because of the varying nature of the coal source and the different processing protocols, different fly ashes exhibit wide ranges of physical and chemical characteristics, resulting in distinct impacts on concrete strength. Thus far, the most adopted method for assessing a given fly ash is specified by ASTM C618, Specification for Coal Fly Ash and Raw or Calcined Natural Pozzolan for Use in Concrete, wherein a series of influential chemical and physical features can be correlated to fly ash's strength activity index (SAI). However, limited knowledge is available on how exactly the individual material attribute affects SAI, so accurately predicting the SAI remains out of reach. Here, we take advantage of recent advances in machine learning to reveal the origins of fly ash's SAI. Leveraging a data set comprising 2,158 fly ash samples, we trained neural network models to predict 28-day SAI based on the sole knowledge of ASTM C618 material attributes. The results demonstrate that SAI is a complex property that does not systematically follow the conventional Class C/F classification. To gain a deeper insight into this matter, we further quantify the influence of each attribute on SAI as captured by the machine learning model.