Why are constructability problems difficult to predict computationally?
Closer collaboration between practitioners from different phases of the building lifecycle can lead to better designs, higher production efficiency, and improved customer or user experience. Such human collaboration allows complex knowledge from subsequent lifecycle phases to be integrated early in the design phase, towards further optimization of the design beyond mechanical parameters. The architecture, engineering and construction (AEC) industry suffers from severe knowledge fragmentation, leading to designs that are rife with so-called buildability/constructability problems when executed on site.
The challenge of disseminating and managing constructability knowledge/documentation lies in its abstract and complex nature. Because of this, it requires expert tacit understandings before improvements can be realized. Consequently, this limits the capacity of explicated or codified constructability guidelines as they are often unable to cover the myriad unique factors in each design and/or construction case. On the other hand, the prediction of specific constructability problems upon design review is readily achievable through high-level human cognition and tacit knowledge embodied in so-called engineering experience/intuition, but lacks scalability.
How we approach constructability knowledge management with artificial intelligence (AI)?
Our method uses Latent Semantic Analysis (LSA), an AI technique, which has been shown to infer patterns and semantic correlations in unstructured textual data comparable to some human cognitive capabilities. Due to the fuzziness and uncertainty inherent in constructability knowledge, LSA is a suitable method.
We start by collecting common textual descriptions of constructability problems and using it as “training” data for the LSA, after some parsing and filtering. The technique enables us to quantify qualitative textual data and infer links between concepts that can be regarded as causes of constructability problems. The latent semantic structure (pattern) that is induced is the basis upon which future constructability problems may be identified computationally - from the correlations inferred between design feature descriptors and associated constructability problems.
What does the algorithm actually produce?
The output is essentially a system of complex semantic links between all the terms of the dataset. These links underpin the ability to predict future issues and gives the impression of “machine understanding” of text. To illustrate, for instance, the algorithm can identify that “slab” is linked to terms like “ceiling” and “void” (building elements); and others like “cure”, “service”, “date”, which suggests slab-related constructability issues. For “basement”, linked terms like “damp”, “drainage”, “flood” are indicators of problems related to “basement”. “Wall” is linked with “stability”, “electric” and “pipe”, “masonry”. These suggest constructability issues regarding walls losing stability due to chasing required for services.
These links are inferred only from the unstructured natural language data, without human tagging. This approach also captures the complex “bias” of the unique conditions of the project factors that cause a constructability problem, without the need to pre-define constructability ontologies or data structures. Additionally, it is automated, thus self-learning as more documentation is produced and inputted.
What are the next steps for developing our solution?
We are looking for pilot partners in AEC industry that can co-develop with us, or even invest in our prototype further, and provide data and real-world test users. We have experience in search engines and complex data analytics, but require the experience of AEC experts to validate the user experience of such an AI-driven constructability management solution, which currently doesn’t exist on the market
