View article

Numerical solutions using machine learning have seen drastic development in recent years, largely due to the advancement in Artificial Intelligence (AI) chips, especially the advent of highly computationally efficient Graphical Processor Units (GPUs) with extensive parallel processing capacities. The application of machine learning-based (ML) methods for solving various structural engineering problems has been explored recently, such as structural analysis and design [1-5], structural health monitoring [6-10], structural optimization [11-16], and so on. The existing research mainly focuses on adopting the ML technique for regression problems. These methods are dataintensive to establish artificial neural networks (ANNs) to solve structural analysis problems through regressions. Among these methods, ANNs are used as a “black box” and the effectiveness highly relies on the qualitative and quantitative of the provided data. However, in structural engineering practice, data is often relatively scarce and costly to generate [17], which causes certain obstacles to adopting the emerging machine-learning techniques in the field of computational structural engineering.

Recently, a novel machine learning method, namely Physics-Informed Neural Networks (PINN), has been proposed by Raissi et al.[18], which is an unsupervised ML technique with the potentials for solving complex mechanical problems. PINN enriches the neural network with information from underlying physical laws, making it possible to train the neural network with few or even no pre-defined datasets while still achieving accurate approximations [19, 20]. PINN incorporates physical …