Design and construction are finally getting an AI upgrade that’s more than shiny demos. In a recent Engineering New Zealand webinar, researchers shared how data-driven methods are improving decisions across the full building lifecycle—design, construction planning, operations, and even deconstruction—using New Zealand’s own experience (including Canterbury earthquake data) as the proving ground.
The opening message was blunt and useful: AI isn’t new, but what’s changed is accessibility and capability. The presenters traced the path from early rule-based systems, through expert systems, to modern machine learning and deep neural networks. Their engineering takeaway was even simpler: success doesn’t come from grabbing the “latest model.” It comes from clearly defining the problem, modeling the knowledge that matters, and then learning from a combination of historical data and domain expertise. In other words: less algorithm worship, more pipeline discipline.
One standout example used earthquake assessment datasets (rapid and detailed evaluations covering thousands of buildings) to train machine-learning models that classify damage states. Random forest models performed especially well in this context, and the team used feature-importance techniques to identify which inputs drive reliable predictions. They extended this beyond “damage” into “functionality,” showing how component-level damage (like roofs and floors, plus non-structural elements such as cladding and glazing) and system-level factors (including foundations and utility supplies) influence whether buildings remain usable after major events. The punchline: resilience isn’t just about the frame—it’s about the whole system.
A second stream of work tackled unstructured data: street-level photos and imagery from damaged buildings. Using deep learning (CNN-style approaches), they aim to detect and cluster damage patterns and even capture “neighbor effects,” where adjacent buildings influence each other’s performance and recovery outcomes.
On the construction side, the team demonstrated a practical planning workflow that connects BIM models (IFC) with documents like standards, regulations, and design reports. Using IDS (Information Delivery Specification) to define machine-readable requirements and large language models to retrieve and validate missing details, their prototype can enrich BIM data and support automated planning with strong test performance.
The final theme was refreshingly honest: the hard part is not AI—it’s data readiness, standardization, and reliability. But if you can solve those, AI becomes a serious co-pilot for faster, safer, more sustainable delivery.