Auto-tune learning framework for prediction of flowability, mechanical properties, and porosity of ultra-high-performance concrete (UHPC)

Machine learning methods are promising to predict key properties of concrete and expedite design of advanced concrete, but the existing methods have limitations in accuracy and generalization performance, because limited dataset size and anomalous data are used to train predictive models. This study presents an auto-tune learning framework for predicting compressive strength, flexural strength, workability, and porosity of ultra-high-performance concrete (UHPC). The presented framework has three features: (1) Structured and unstructured data are combined. (2) Anomalies and inappropriate variables in the dataset are identified and removed using an unsupervised anomaly detection method based on isolation forest and combined mutual information and univariate linear regression. (3) The hyperparameters of machine learning models are optimized using tree-structured Parzen estimator with k-fold cross-validation. Auto-tune predictive models are developed by integrating the presented learning framework and Light Gradient Boosting Machine (LightGBM). The results showed that the developed method achieved high prediction accuracy. The auto-tune models are used to study the effects of mixture design variables on the properties. This research will greatly promote material development by reducing experiments.