Figure 4
The diagram shows a complete processing pipeline that begins with four inputs aligned vertically on the left: K subscript P R, M subscript P R, f subscript P R, and a general group labeled Hyperpar, representing hyperparameters. Below the hyperparameter input, two neural network modules are illustrated. The upper module starts with an input of zero, passes through a fully connected dense layer, then through a hyperbolic tangent activation function, and outputs k subscript S E. The lower module also begins with a zero input, passes through a dense layer, followed by a sigmoid activation function, and produces m subscript S E. These outputs, along with the other four input parameters, are fed into a large dashed rectangular box to the right, which represents the computational pipeline. The steps inside this box are given below. Build K subscript S E, M subscript S E from lowercase k subscript S E and m subscript S E, scale, rotate. Assemble K tilde, M tilde. Adjust f of (t) to f tilde of (t). C tilde equals alpha M tilde plus beta K tilde. Newmark Beta: K tilde u tilde of (t) plus C tilde u dot tilde of (t) plus M tilde u double dot tilde of (t) equals f tilde of (t). Reduce u tilde of (t) to N subscript P R. The output from the box is the time-dependent displacement vector u tilde of (t), shown on the far right side.Architecture of the informed ANN for transient simulations. Source: Authors’ own work