Transforming Building Design Through Generative AI: ZURU Tech’s Dreamcatcher

Introduction to ZURU Tech’s Vision

The Role of Generative AI in Building Design

Enhancing Design with LLMs

Understanding the Challenge of Floor Plan Generation

Improving Results Through Innovative Approaches

Dataset Preparation for Effective Model Training

Advanced Prompt Engineering Techniques

Workflow Architecture for AI Optimization

Fine-Tuning Methodologies for Enhanced Performance

Evaluation Framework for Assessing Outcomes

Measuring Success: Results and Impacts

Conclusion and Future Directions in Building Design Innovation

About the Authors

Transforming Construction with ZURU Tech and Generative AI

ZURU Tech is pioneering a revolution in the construction industry, aiming to transform how we design and build everything from townhouses to hospitals, office towers, schools, and apartment blocks. With the launch of Dreamcatcher, a user-friendly platform, ZURU enables users—regardless of their experience level—to collaborate seamlessly in the building design and construction process. With just a click, an entire building can be ordered, manufactured, and delivered right to the construction site for assembly.

Enhancing Creativity with Generative AI

In collaboration with the AWS Generative AI Innovation Center and AWS Professional Services, ZURU has developed a sophisticated, text-to-floor plan generator that leverages the capabilities of generative AI. This tool allows users to describe their desired building in simple, natural language. For example, a user can request, “Create a house with three bedrooms, two bathrooms, and an outdoor space for entertainment,” and the system will generate a tailored floor plan within a 3D design space. This is particularly empowering for individuals without a technical background in architecture and construction, providing them a pathway to create well-structured living spaces.

The Challenge of Floor Plan Generation

Successfully generating a house plan in Dreamcatcher’s 3D building system first necessitates the ability to provide a 2D floor plan based on user prompts. ZURU identified two key criteria for this generation:

1. Understanding Room Functionality: The model must accurately grasp the purpose of each room and their configurations within a two-dimensional vector system.
2. Mathematical Integrity: Each room’s dimensions and spatial relationships must adhere to specified criteria.

To facilitate rapid Research & Development, ZURU crafted a novel evaluation framework measuring outputs against these criteria, which ultimately guided their experimentation with generative adversarial networks (GANs) and LLMs (Large Language Models).

Improving Outcomes with Advanced Models

To enhance the results obtained from an initial GPT2 LLM, ZURU participated in two key experiments: Prompt Engineering and Fine-Tuning.

• Prompt Engineering: Utilizing Anthropic’s Claude 3.5 Sonnet in Amazon Bedrock, the ZURU team introduced contextual examples into prompts, optimizing the model’s understanding and reactions to user inputs.
• Fine-Tuning: The team employed variants of the Llama 3B model, adjusting model weights using high-quality examples to improve output accuracy.

Dataset Preparation and Quality Control

The first step involved aggregating thousands of floor plans from publicly available resources, meticulously reviewed by in-house architects. ZURU developed a streamlined review application, enabling quick decision-making to either approve or reject plans based on compatibility with their building system. A rigorous dataset preparation effort filtered out low-quality data, significantly enhancing the subsequent fine-tuning and prompt engineering processes.

Dynamic and Efficient Prompt Engineering

ZURU’s approach to prompt engineering included dynamic matching for few-shot prompting and prompt decomposition, dramatically improving the quality of generated floor plans.

• Dynamic Few-Shot Prompting: This process retrieves relevant examples in real time, enhancing the model’s context and facilitating improved performance.
• Prompt Decomposition: Complex tasks are broken down into simpler components, allowing each to be optimized individually, thus enhancing the overall results.

Innovative Workflow Architecture

The architecture used for prototyping incorporates several steps to optimize the AI model’s performance based on user queries. Here’s a simplified breakdown:

1. Feature Identification: The system pinpoints unique features of the requested house to search for relevant examples.
2. Example Retrieval: Utilizes Amazon Bedrock and OpenSearch as a vector database for efficient data retrieval.
3. Floor Plan Generation: Relevant examples are fed into the AI model, resulting in the creation of a new floor plan.
4. Reflection and Verification: The AI model self-assesses its output based on user requirements, ensuring adherence to specified criteria.

Fine-Tuning for Optimal Accuracy

Two methods were explored for optimizing LLMs for floor plan generation: Full Parameter Fine-Tuning and Low-Rank Adaptation (LoRA). The latter proved to be a more computationally efficient approach, enhancing time and resource management without compromising significant performance.

Insights from the Evaluation Framework

To gauge the effectiveness of various techniques, ZURU deployed a comprehensive evaluation framework focusing on two main metrics: adherence to user instructions and mathematical correctness of the generated floor plans. Results showcased a staggering 109% increase in instruction adherence quality with the prompt engineering approach, while full fine-tuning exhibited a 54% improvement in mathematical accuracy.

Conclusion

ZURU Tech’s commitment to transforming the design and construction landscape is evident in the development of advanced tools like the text-to-floor plan generator. By integrating cutting-edge technologies in generative AI and focusing on user experience, ZURU not only streamlines the construction process but also democratizes access to architectural design.

For those eager to delve deeper into ZURU’s partnership with AWS and explore the transformative potential of generative AI in construction, we invite you to reach out to your account team.

About the Authors

Federico Di Mattia, Niro Amerasinghe, Haofei Feng, Sheldon Liu, Xuefeng Liu, Simone Bartoli, Marco Venturelli, Stefano Pellegrini, and Enrico Petrucci contribute diverse expertise across AI, architecture, and cloud solutions to propel ZURU Tech’s innovative mission.