Hyperparameter Tuning for Artificial Neural Network Pre dicting Concrete Compressive Strength
S.H. Seuntjens (TU Delft - Civil Engineering & Geosciences)
B. Šavija – Mentor (TU Delft - Materials and Environment)
Zhi Wang – Mentor (Tianjin University)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Concrete compressive strength is the most frequently used and most important mechanical propertyof concrete. National and international building codes (such as the Eurocode) frequently use the compressive strength for design with concrete.
In some cases, instead of testing the concrete specimens under compressive loading in laboratories, predicting the compressive strength using machine learning predictions could be a good alternative. The ingredients making up the concrete mix and the curing age can be used as predictors for the compressive strenght.
Artificial Neural Networks (ANNs) are machine learning algorithms that have been used since the nineteen sixties and were inspired by the way neurons work in the brain. Previous researches suggest that ANNs have great potential to predict concrete compressive strength.
In this research, an Artificial Neural Network was set up using the Keras framework and was trained with concrete composition data consisting of examples of concrete recipes and their respective concrete compressive strenghts. Then, three hyperparameter optimization methods (for loop, grid search and Bayesian optimization) were implemented in several runs. The resulting hyperparameters were used to create ANNs. Afterwards, the three methods were compared with respect to several met rics (Rsquared score, root mean square error and running time) to see which one is the relatively best method to provide hyperparameters for the predefined ANN that learns from concrete composition data.
The best runs of the three hyperparameter optimization methods show similar goodnessoffit (with negligible difference). Among these best runs, grid search has the shortest running time. Bayesian optimization provides the highest Rsquared score, and has root mean square error of 4.45 [MPa] and a reasonable running time of 73 minutes. Therefore, Bayesian optimization is considered the preferable hyperparameter optimization algorithm for the concrete compressive data used as goodnessoffit is in most cases to be considered the decisive metric for this application. Further research trying runs with larger parameter grids and spaces, or using other tuning methods could result in even better goodness offit results.