BRITE Relaunch: Improving Structural Health by Advancing Interpretable Machine Learning for Nonlinear Dynamics

This Boosting Research Ideas for Transformative and Equitable Advances in Engineering (BRITE) Relaunch award will focus on advancing interpretable machine learning to match modeling accuracy with transparency. This will provide structural engineers with a superior and trustworthy tool to model nonlinear dynamical systems. Modeling the complex behaviors of structures and materials under various types of dynamic loads remains a major challenge for smart structures, structural control, nonlinear system identification, damage detection, and earthquake engineering. Machine learning is becoming a popular approach for meeting this challenge. However, there is a significant gap between knowledge of the physics of these systems, the engineering practice design of these systems, and the models produced from machine learning methods. Machine learning lacks interpretability and transparency. This research project will develop systematic solutions with reasoning based on the engineers’ knowledge and training, physics-informed, and empowering the engineers’ judgment. This research directly benefits society in terms of improving infrastructure health, mitigating the consequences of earthquake, wind hazards, and climate change. This research project builds upon the project leader’s past work to advance “nonlinear static function approximation using interpretable machine learning”, given its direct use in approximating nonlinear constitutive relations and its use in approximating nonlinear integrands in ordinary differential equations for nonlinear dynamics. To achieve interpretable and physics-informed machine learning methods, this research project will create new algorithms and implementation procedures. Neuromanifold theories in advanced applied mathematics will be employed to make the training process of sigmoidal neural networks interpretable. Graph theory will be leveraged to create knowledge graphs so that nonlinear static function approximation using interpretable machine learning can be automated during initialization to approximate nonlinear static functions and can be used for deep learning. In addition to extensive cross-validations, a major application of the project's approach will be investigated by using real-world data in a digital twin setting, the state-of-the-art system-level modeling framework in structural engineering. Also, a comprehensive laboratory demonstration and validation will be carried out using timber beam-column joints to generate broad interest in the broad relevance of nonlinear dynamics in structural engineering. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.