Deep learning-based surrogate model to estimate scour around slab-on-grade foundations subjected to flooding conditions

Residential foundations in coastal areas are susceptible to loss of soil support due to erosion and scour during storm surge events. The current methods to address these issues are largely based on empirical evidence and lack sensitivity to site-specific conditions. Thus, a research study was designed and performed to develop a computationally efficient deep learning-based surrogate model to estimate scour around slab-on-grade foundations on sandy soils over a wide range of flooding conditions. A fully coupled 3D numerical model, calibrated using published laboratory test studies, was used to simulate scour around slab-on-grade foundations. This model was used in conjunction with a space-filling experimental design to generate sample points required to develop a surrogate model. Two deep neural network (DNN)-based surrogate models were developed and tested for predicting scour for varying flow and soil conditions. The DNN-based surrogate model, augmented with modified loss function to penalize un-physical predictions, was shown to achieve the best performance among the studied methods. Monotonicity analysis on the model showed that flood velocity, duration, and grain size properties of soils were found to have a major influence on the predicted maximum scour depth as compared to flood depth. The surrogate model developed can serve as an aid in achieving improved storm surge hazard preparedness, as well as proactive planning for post-hazard recovery efforts in vulnerable coastal communities.