Figure 4
The flowchart shows stages in surrogate model development. An optional step labelled variable screening, if inexpensive, leads to data sampling. Data sampling includes grid search, random search, Latin hypercube sampling, quasi random low discrepancy sequences, and adaptive sampling. The process continues to selection of neural network model, listing feedforward network F F N, physics informed neural network P I N N, long short term memory network L S T M, and transformer. This connects to selection of activation function, including sigmoid, tanh, Re L U, softmax, softplus, and swish. The next step is tuning, including hyperband, Bayesian optimisation, and neural architecture search. The flow proceeds to evaluation, which checks accuracy using coefficient of determination, root mean square error, and mean absolute percentage error. A decision follows. If yes, the process moves to applying surrogate model. If no, it loops back to neural network selection. A legend indicates process, optional process, and decision symbols.Flowchart showing the steps required for developing a surrogate model