TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 19—21 · FACTUAL · DURABLE
The surrogate model
A surrogate outside its training range still answers just as quickly, and is not close.
The idea
A surrogate is a model trained to approximate something slower or more expensive to compute. In architecture that is usually a physics simulation — daylight, energy, wind, heat — though the original can be an experiment or any costly calculation. The simulation is run many times to make training data. The surrogate learns the mapping from inputs to results, then answers in seconds what the original takes far longer to compute. The catch is where it learned. Inside the range of the training runs it is usually close. Outside that range it still answers, just as quickly, and is not close.
Why it matters
A surrogate makes performance questions fast enough to ask at sketch stage, which is when they matter. It also makes a wrong number fast enough to put on a jury sheet.
See it in the studio
A daylight surrogate trained on rectangular rooms with side windows gives a crisp daylight factor for your courtyard scheme with a clerestory. The number is confident. The courtyard was never in its training set. Nothing in the output says so.
Watch for this
Quoting a surrogate's figure without its range. "Daylight factor 2.8%" means nothing on its own. "2.8%, from a surrogate trained on side-lit rooms, and our scheme is top-lit" means something. What it means is: run the real simulation.
Try it
Find any fast performance tool you use — a daylight or energy quick-check. Find out what it was trained or calibrated on. If you cannot find out, write that down as the answer.
Prove it
Explain what a surrogate imitates and why it is faster than the real simulation, and name the validity limit that must be stated whenever a surrogate's answer is used.
How it works
Surrogates are ordinary regression models — neural networks, gradient-boosted trees — trained on simulation outputs. Good practice builds in an applicability-domain check that flags inputs outside the training range and reports uncertainty with each prediction. Recent building-performance work does exactly this. FIND OR INVENT on Foundations is the related idea: a surrogate predicts, it does not look up.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.