| description abstract | This paper introduces CLEMSON, an automated machine-learning (AutoML) virtual assistant (VA) that enables engineers to carry acCeLErated, siMulation-free, tranSparent, reduced-Order, and infereNce-based fire resistance analysis with ease. This VA learns from physical observations taken from real fire tests to bypass bottlenecks and ab initio calculations associated with traditional structural fire engineering methods. CLEMSON leverages a competitive ML algorithm search to identify those most suited for a given problem and then blend them into a cohesive ensemble to realize faster and reduced-order assimilation of predictions, thereby attaining higher accuracy and reliability. In addition, this VA is designed to be transparent and hence is supplemented with explainability measures to allow users to identify key factors driving its rationale and predictions. Once fully realized, CLEMSON augments its inner workings into a graphical user interface that can be used in a coding-free manner and with enriched visualization tools to allow users to directly harness the power of ML without the need for special software. To showcase the merit of the proposed VA, CLEMSON is applied to assess classification and regression problems by means of evaluating fire resistance rating, as well as temperature rise history and deformation history of concrete-filled steel tubular (CFST) columns via five algorithms, namely: extreme gradient boosted trees, light gradient boosted trees, neural networks, random forest, and TensorFlow. Finally, this work also introduces three new and functional performance metrics that are explicitly derived for structural fire engineering applications and hence can be used to cross-check the validity of ML models. | |
