StructuresNet and FireNet: Benchmarking databases and machine learning algorithms in structural and fire engineering domains

Abstract

Machine learning (ML) continues to rise as an effective and affordable method of tackling engineering problems. Unlike other disciplines, the integration of ML into structural and fire engineering domains remains deficient. This is due in part to the lack of benchmark databases to compare the effectiveness of ML models. In order to bridge this knowledge gap, this paper presents a benchmark examination of common supervised learning ML algorithms that can be easily deployed into structural and fire engineering problems. The selected algorithms include; Decision Trees (DT), Random Forest (RF), Extreme Gradient Boosted Trees (ExGBT), Light Gradient Boosted Trees (LGBT), TensorFlow Deep Learning (TFDL), and Keras Deep Residual Neural Network (KDP), and are used with their default values to establish a proper benchmark against six databases. The compiled datasets have been thoroughly tested and span two domains, structural engineering; 1) elemental response of concrete-filled steel tubular (CFST) circular columns at ambient conditions, 2) shear response of cold-formed steel (CFS) channels with slotted webs, 3) compressive strength of concrete, 4) fatigue life data, 5) shear strength of reinforced concrete (RC) beams and FRP-strengthened RC beams; and fire engineering, 6) fire behavior of RC concrete columns in terms of spalling occurrence and fire resistance. This study also investigates a variety of commonly used performance metrics that are applicable to regression and classification-based ML problems. We invite ML users to apply their models to the presented databases to establish a benchmark by mean of external validation and then extend their models to other problems and databases. Collectively, the presented work establishes the first step towards a unified framework that can be used to accelerate the adoption of ML into structural and fire engineering domains.