In recent months, generative AI tools such as OpenAI’s ChatGPT have exploded in use across a number of industries, including law, education and advertising + marketing. Generative AI is also beginning to make its presence known in the Built World. Applications range from leveraging generalist solutions – such as chatbots to optimise property management, image generators (such as DALL-E and Midjourney) and enhancers (such as LetsEnhance and Oda Studio) to augment real estate sales and synthetic data generators (such as Mostly.AI, MindTech and LatticeFlow) to train AI models – through to specialist products tailored to Built World use cases.

Generative Models for the Built World

Part of the wider AI ecosystem, generative training models learn from large datasets to create new, similar data. Core applications of generative models in the built world focus on generative adversarial networks (GANs), dense pre-trained networks (DPT) and auto encoders, which can be used to synthesise and augment data, as well as automate core aspects of design workflows.

Data Synthesis

Generative model outputs data in a different format to the input, with the results able to be used either qualitatively or quantitatively for Built World workflows.

• Map conversion and creation (GIS)
• Climate data estimation and simulation
• 3D reconstruction
• Environmental performance simulation
• Cross view image generation

Data Augmentation

Generative model outputs data in a similar format to the input data, but with improvements in resolution or completion.

• Digital surface model (DSM) enhancement
• Satellite image and street view enhancement
• Data upsampling
• Image enhancement and restoration
• Discovery of low carbon concrete formulations

Design Automation

Generative model outputs data in a different format to the input, with results able to be used qualitatively to inspire or automate certain workflow processes.

• 3D model generation
• Design analysis, assistance and generation
• Floor plan generation and optimisation
• Custom and high definition 3D renders
• Text to Building Information Model (BIM)

Commercialising Gen-AI Applications

• There are currently a handful of commercialised applications in the built world, and we expect to see a number of new applications come to market in the coming years. At present, there are a growing number of pre-seed startups leveraging generative models to optimise building design. Though the technology is still in its infancy, the potential for future market development is notable.
• Functions include instant 3D model creation from a 2D image, custom material render generation (such as Poly), optimised floor plan layouts (such as Maket and A-Space), and even the generation of a Building Information Model (BIM) from simple descriptive text (such as Hypar).
• Applications to building design are promising, yet there remains a long way to go before solutions can offer efficiency gains similar to that of which ChatGPT is currently enabling across other industries. This is because, unlike readily available image or text data, there is a shortage of building design and construction data.

Digitisation is Key to Resolving Data Shortages

• Data shortages are only exacerbated by the highly heterogenous and context-specific nature of building-level data, which makes it much more difficult to collect a similar set of data. For data types gathered at much larger scales - such as satellite imagery and GPS data – training datasets are large enough to harness the full power of generative models. This can be seen in solutions such as , which leverages generative models to create a digital twin of the earth.
• While generative AI presents significant potential, in the short term digitisation and automation will be key – especially for the industry’s most analogue aspects, such as construction materials procurement and in-person energy audits.
• Startups such as Satellite Vu plan to collect the data that fills the gap between the high-fidelity drone or street level data and general large scale satellite mapping data. Satellite Vu’s CTO Tobias Reinicke explains: “With our high resolution thermal imagery, we can support decision makers in evaluating the thermal characteristic of building structures – enabling inferences to be made of their efficacy. Satellite Vu is already working with a multitude of data streams which, when put together, can be trained to identify thermal anomalies and outliers that require intervention and, with the hyper-stereo imagery capability, can accurately map thermal signatures in three dimensions”.
• For other applications, it will be essential to look beyond the current hype and frenzy and consider whether a similar output can be delivered using a combination of other AI technologies. This could be important as, while generative models might be able to derive slightly better results, higher computation costs and the associated carbon intensity of training models might make it difficult to justify large scale deployment.

Startups mentioned: Maket, A-Space, Hypar, Satellite Vu, Poly, Blackshark.ai

‍

‍

‍