A Review of Physics-based Machine Learning in Civil Engineering

ML and DL, e.g., deep neural networks (DNNs), are becoming increasingly prevalent in the scientific process, replacing traditional statistical methods and mechanistic models in various commercial applications and fields, including education [1], natural science [2, 3] medical [4-6] engineering [7-9], and social science[10]. ML is also applied in civil engineering, where mechanistic models have traditionally dominated [11-14]. Despite its wide adoption, researchers and other end users often criticize ML methods as a "black box," meaning they are thought to take inputs and provide outputs but not yield physically interpretable information to the user[15]. As a result, some scientists have developed physics-based ML to reckon with widespread concern about the opacity of black-box models [16-19]. The civil engineering ML models are created directly from data by an algorithm; even researchers who design them cannot understand how variables are combined to make predictions. Even with a list of input variables, black-box predictive ML models can be such complex functions that no researchers can understand how the variables are connected to arrive at a final prediction.