From Waste to Structure: A Deep Reinforcement Learning Approach to Circular Design
Name
Soerensen-karljo-smarchs-smcee-arch-cee-2024-thesis.pdf
Description
Thesis PDF
Size
26.21 MB
Format
Adobe PDF
Checksum (MD5)
b419537c34bea69bf7e6fc0b2b643c9a
Author(s)
Sørensen, Karl-Johan I.
Advisor(s)
Mueller, Caitlin T.
Sass, Lawrence
Date Issued
May 2024
Publisher
Massachusetts Institute of Technology
Abstract
The design-to-construction process of buildings predominantly follows a top-down linear workflow, where a design is drawn and subsequently refined to determine the required materials and components. This approach assumes an infinite material supply or the capability to manufacture what is needed for the design. Constructing in this manner is resource-intensive and wasteful, making it incompatible with our global climate goals. One way to significantly reduce our material and environmental footprint is by extending the lifespan of building materials through circular design practices. In this approach, the available materials define the architecture, inverting the process from top-down to bottom-up. This method, known as Inventory-Constrained Design, enables the creation of new buildings using materials sourced from construction and demolition waste streams. These inventories, characterized by their non-standard and uniquely varied elements, are hard to design with due to the enormous quantity of possible combinations of even a few discrete elements. Identifying a feasible design that aligns with the designer's intent and meets functional requirements becomes an overwhelmingly time-consuming task, heavily reliant on manual trial and error. Computational optimization has been implemented to automate the process, but state-of-the-art algorithms still require manually pre-defining a parametric target design-space or take too long to compute when applied to larger problems.
This thesis proposes a new method for circular design utilizing Deep Reinforcement Learning (RL) to design structures, requiring only a design gesture and the inventory as input. It works by training an artificial neural network to sequentially assemble a structure from inventory elements, following the gesture while meeting a structural goal. Hence, the design layout directly arises from available inventory. After training, the neural net can be employed instantaneously to design new structures with new inventories without any significant computational expense. To evaluate the effectiveness of the RL method, it is applied to the specific problem of inventory-constrained design of planar roof trusses and demonstrated in a realistic example of assembling a long-span roof from a disassembled transmission tower.
MIT Department
Massachusetts Institute of Technology. Department of Architecture
Massachusetts Institute of Technology. Department of Civil and Environmental Engineering
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
