How machine learning models trained on engineering data are compressing weeks of design iteration into a single afternoon.
The traditional MEP design process is sequential and slow. A mechanical engineer sizes the HVAC system. An electrical engineer sizes the distribution boards. A plumbing engineer designs the wet services. Each discipline works from the previous discipline's outputs — a process that takes weeks and accumulates errors at every handoff.
AI-assisted design changes the model. Instead of a linear sequence, we run a generative model that evaluates thousands of design configurations simultaneously — weighting energy efficiency, capital cost, spatial coordination and maintenance access against a project-specific criteria matrix. The result is not a single design, but a Pareto-optimal frontier of solutions, each representing a different balance of competing priorities.
How the model works
Our model has been trained on twelve years of completed MEP projects, each tagged with actual post-occupancy energy performance, construction cost outturns and maintenance records. The training data captures not just what was designed, but how each design performed in practice — a crucial distinction that allows the model to identify configurations that look good on paper but underperform in operation.
"The model surfaces what experience would have taken years to learn — which design choices actually deliver in practice."
Given a new project brief — building type, floor area, occupancy profile, climate zone, energy target — the model generates an initial solution set within four hours. An experienced engineer then reviews the frontier, selects candidate designs and applies site-specific constraints that the model cannot yet infer from structured data alone.
What it changes for clients
- —Design options are presented at Stage 2, not Stage 4 — earlier client decisions, less design abortive
- —Energy modelling is embedded in the design process, not bolted on at the end
- —Capital cost comparisons between systems are available from day one
- —Coordination clashes between disciplines are caught before detailed design begins
The time saving is real but secondary. The more significant impact is on design quality. When an engineer can evaluate a hundred configurations in the time it previously took to evaluate three, the final design is simply better — more carefully optimised, more thoroughly stress-tested, and more likely to meet its energy target in operation.
The limits of the approach
AI-assisted design is not autonomous design. The model cannot account for client preference, political constraints, supplier relationships or the thousand small judgments that an experienced engineer makes in the course of a project. It is a tool for expanding the range of options that a human engineer can consider — not a replacement for the engineering judgment required to choose between them.
