TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Studio Practice · AGE 18—21 · FACTUAL · EVOLVING
Environmental analysis with AI
Screen fast with a surrogate, and confirm properly before the number drives a decision.
When to use
Early, when you want to rank options by heating, cooling, daylight or shading before a full model exists — and every time a fast number is about to drive a decision.
The method
One: name the question and the decision it feeds. Two: get the inputs right — orientation, glazing area, envelope, and a climate file for your city, not a default one. Three: use the fast tool — a surrogate model, or a script the language model helped you write — to screen and rank, not to size: a surrogate sizes nothing, and decides no compliance figure, until it has been validated against a full simulation for your kind of building. Four: ask what the tool was trained on and whether your building sits inside that range; a surrogate is only valid near the cases it learnt from. Five: confirm the chosen option with a full simulation or an engineer before it becomes a design decision. Screen fast, confirm properly.
Watch for this
A surrogate's accuracy score describing its own training set, not your building. A 2026 workflow paper reports fit scores above 0.97 on a benchmark of 768 simulated configurations and then says plainly that these must be read as interpolation within the benchmark, not as evidence about real buildings or plant sizing. Read every accuracy claim that way.
Try it
Ask a language model for the cooling load of a simple room you describe. Then ask it which climate, which construction and which dataset its number assumes. Note whether it can say. Then find a climate file for your own city and see how many of those assumptions were wrong.
Prove it
For one early-stage screening, show the inputs you gave, the validity limits the tool stated, and the confirmation step before the number drove a decision.
How it works
A surrogate is a learning model trained on the outputs of a physics simulation, so that it can answer in milliseconds what the simulation answers in minutes. The 2021 daylight framework on this card trained a network on 2,880 simulations of a single-window shoebox room; the 2026 heating-and-cooling workflow trained on 768 simulated configurations with explicit validity gates and uncertainty bands. Both are useful for ranking; neither knows a courtyard in Hubballi unless one was in the training set. India has climate files for many cities, and the tools that accept them are changing quickly. That, with the pace of new surrogate research, is why this card is marked EVOLVING.
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.