Estimation of wind pressure field on low-rise buildings based on a novel conditional neural network

Abstract

The estimation of wind pressure on building envelopes is critical for accurate risk and resilience evaluation of coastal structures. It is possible and convenient sometimes to interpolate existing wind-tunnel tests for the estimation of untested buildings. However, the complex bluff-body aerodynamics pose a challenge for wind tunnel test interpolation. Specifically, the nonlinear interpolation needs to be conducted simultaneously on coordinate and wind and building condition, which cannot be accomplished using conventional interpolation methods. To address this issue, a novel machine learning approach called Conditional Neural Network (CondNN) is developed. In the CondNN, a two-level architecture is utilized to interpolate existing test data for a certain building (i.e., the lower level takes the coordinate on building envelope as input) and across different buildings (i.e., the upper level takes the wind and building condition as input). Wind tunnel measurements are further used for verifying the performance of the proposed conditional natural network for wind pressure estimation. The CondNN not only paves a path toward the fast fragility evaluation of a diverse set of buildings exposed to coastal wind hazards, but also illuminates the use of machine learning techniques in a broader spectrum of engineering applications.