Sequential Machine Learning Model: Comparison of Optimizers for Concrete Compressive Strength Prediction

Abstract

Concrete is the most widely used construction material today due to its exceptional ability to withstand compressive forces, commonly referred to as CS. Determining the value of CS for concrete involves conducting various tests, with the uniaxial compression test on concrete specimens being the most used, assessing strength at different time intervals. This research focuses on predicting the value of concrete's CS at 28 days using 16 predictive models based on artificial neural networks, each employing different optimizers, including Adam, Adamax, Rmsprop, and Nadam. Input data includes aggregate properties, cement type, and the proportions of its components, such as water, cement, and aggregates. The neural networks were built using Google's TensorFlow, with two hidden layers that vary in the number of neurons in the first hidden layer (13, 16, 19, 22) and 8 neurons in the second hidden layer. Training was conducted over 400 epochs.The comparison of optimizers reveals that Nadam exhibits the best performance in both the training and prediction stages, with R² values of 0.83 and 0.87 respectively. The increase in the number of input variables in the prediction results in higher accuracy in predicting CS.