Data-driven approximation algorithms for rapid performance evaluation and optimization of civil structures

Abstract

This paper explores the use of data-driven approximation algorithms, often called surrogate modelling, in the earlystage design of structures. The use of surrogate models to rapidly evaluate design performance can lead to a more indepth exploration of a design space reduce computational time of optimization algorithms. While this approach has been widely developed and used in related disciplines such as aerospace engineering, there are few examples of its application in civil engineering. This paper focuses on the general use of surrogate modelling in the design of civil structures, and examines six model types that span a wide range of characteristics. Original contributions include novel metrics and visualization techniques for understanding model error, and a new robustness framework that accounts for variability in model comparison. These concepts are applied to a multi-objective case study of an airport terminal design that considers both structural material volume and operational energy consumption.

Design exploration of quantitative performance and geometry typology for indoor arena based on self-organizing map and multi-layered perceptron neural network

“…A surrogate model can approximate a high-fidelity but time-consuming function in reasonable accuracy, based on sampled/labelled data obtained by design of experiments (DoE) of the high-fidelity function [15,25]. This method has been applied in building designs [13][14][15][16].…”

“…Some of the methods have been used in engineering or architectural designs, including poly-nominal regression and response surface method (RSM) [16,[27][28][29], multi-layer perceptron neural network (MLPNN) [15,29] [30,31], random forest (RF) [15,32,33], radial basis function network (RBFN) [14,15], kriging [15,34,35]. In [15], these methods are used to support surrogate models for a long-span building design focusing on structural self-weight and energy. The results in [15] indicated that the MLPNN has the fastest speed and smallest errors in the data approximation of structural weight and energy consumption for the design example.…”

“…In [15], these methods are used to support surrogate models for a long-span building design focusing on structural self-weight and energy. The results in [15] indicated that the MLPNN has the fastest speed and smallest errors in the data approximation of structural weight and energy consumption for the design example.…”

“…Mean squared errors (MSE) is used as the cost/error function which should be minimized during the training process. This MLPNN model is also used in [15] for a long-span building to predict structural self-weight and energy consumption. To overcome the problems related to overfitting and generalization, a validation and test process is used.…”

“…In this light, an efficient design method is needed to satisfy the aforementioned demands, therefore, to support the design exploration of indoor arenas, which is also the motivation of this paper. Nowadays, several computational design methods have been used to support design exploration of architectural conceptual design, including multiobjective optimization (MOO) [3][4][5][6][7][8][9][10][11][12], surrogate model based on supervised learning [13][14][15][16], and unsupervised clustering based on selforganizing map (SOM) [17][18][19]. Fig.…”