Parametric Modelling and Generative Design - A Multi-Step Machine Learning Approach for Design and Optimization of Network Tied-Arch Bridges
Author: Sophia Kuhn
Language: English
Abstract
Conceptual structural design today relies heavily on the intuition and experience of the structural engineer, often includes an investigation of similar reference projects, and is mostly a time-consuming and demanding task that is characterized by many iteration steps. This thesis hypothesizes that conceptual structural design as we know it today can be greatly assisted and improved by using modern machine learning (ML) methods. By developing a multi-step ML approach to improve the design and optimization process of network tied-arch bridges, the author shows the wide potential of applying ML algorithms to the conceptional structural design phase.
Based on a pre-processed data set of already built network tied-arch bridges worldwide, a cluster analysis is performed to find a structure within the data. The two unsupervised learning algorithms KPrototype and DBSCAN were successfully applied to the data set containing a mix of continuous and categorical bridge parameters. A consistent structure was identified by both algorithms showing strong similarities of all bridge parameters for network tied-arch bridges that are within similar span ranges.
Subsequently, a model is trained based on the available prior data that is capable of predicting suitable bridge parameters for a new project situation in a predefined order. Due to its capability to handle mixed data types, the open-source CatBoost library is chosen and applied to implement regression and classification. It applies the supervised learning method of gradient boosted decision trees. All programming-related tasks have been implemented in Python. Based on the predictions made by the prior data model, a parametric structural model is constructed using Grasshopper and Karamba3D in the Rhino 7 environment. This Generative Model provides real-time structural analysis feedback for any parameter changes. Furthermore, the plug-in Octopus is applied to conduct multi-objective optimizations of the bridge parameters using the genetic optimization algorithm HypE. Objective functions calculating material costs, structural efficiency, and aesthetic quality are developed.
The use of machine learning algorithms to generate new bridge designs is proven to reduce optimization time to find near optimum design solutions. The prior data model is shown to provide well advised parameter predictions at the start of the project. The predictions are based on detailed investigations of completed construction projects and therefore show a great potential to improve future construction projects by providing easy access to more complete information at an early stage. An identified limitation of the implemented approach is the small size of the data set the model is trained on. Additionally, the input of expertise and creativity by the structural engineer remain necessary as the machine learning model can adopt potential mistakes from the past and does not promote structural innovation.
