Locally Connected Neural Networks as Surrogate Models for Stochastic Analysis with Spatial Variability

Using machine-learning models as surrogate models is a popular technique to increase the computational efficiency of stochastic analysis. In this technique, a smaller number of numerical simulations are conducted for a case, and obtained results are used to train machine-learning surrogate models specific for this case. This study presents a new framework using deep learning, where models are trained with a big dataset covering any soil properties, spatial variabilities, or load conditions encountered in practice. These models are very accurate for new data without re-training. So, the small number of numerical simulations and training process are not needed anymore, which further increases efficiency. The prediction of bearing capacity of shallow strip footings is taken as an example. More than 12000 data are used in training. It is shown that one-hidden-layer fully connected networks are ineffective for complex problems, where deep neural networks show a competitive edge, and a deep-learning model achieves a very high accuracy (the root-mean-square relative error is 3.1% for unseen data).

Keywords: bearing capacity, locally connected neural networks, surrogate models, stochastic analysis