TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Studio Practice · AGE 19—21 · POSITIONAL · METHOD
A surrogate answer needs a real simulation behind it
A surrogate number never leaves the studio without a real simulation behind it.
When to use
Any time a fast performance number — daylight, heating, cooling, wind — is about to move from your screen into a drawing, a jury sheet, a client report or a submission.
The method
Three stages and a rule. Explore: use the surrogate — the fast model trained to imitate a simulation — to rank many options in minutes; this is what it is for. Defend: take the two or three options you would actually argue for and run them through a physics engine with your site's weather file — EnergyPlus, Radiance, a CFD run, or an engineer's model. Re-check: for those three, put the surrogate's figure and the simulation's figure side by side. The gap between them is the surrogate's error on your building, and it tells you whether the ranking from the explore stage can be trusted. Then check the simulation itself: change the two inputs you are least sure of — the occupancy, the infiltration, the glazing g-value — and see how far the answer moves. That range is the number's honest width, and it goes on the sheet beside the number. The rule: a number that leaves the studio carries the simulation's name, its inputs and its date — never the surrogate's alone. If the simulation has not been run, the sheet says "indicative, not yet simulated."
Watch for this
A surrogate figure that survives into a later drawing because nobody tracked which numbers were fast and which were real. Mark them differently from the first day — a colour, a suffix — so the provenance is on the sheet, not in your memory.
The Lab's note
This is the Lab's method. It is a rule about what may leave the room, not a view on which tool is better. Surrogates are good; we use them. But a number a client will act on needs physics behind it, because a surrogate is just as confident outside the range it learnt from as inside it, and it cannot tell you which is which. The physics does not make the number true — the weather file, the boundary conditions, the occupancy, the material properties and the mesh can all be wrong — but it makes the number checkable, input by input, which a surrogate is not. A surrogate figure standing alone is the first thing this strand tells a student not to put in a submission.
Try it
Take one scheme with three massing options. Rank them with any fast daylight or energy tool. Then run the top one through a full simulation with your city's weather file. Compare the two figures and write one sentence on what the difference tells you about the other two. Then run the simulation twice more with your two least certain inputs moved to their plausible extremes, and report the result as a range, not a number — "between 41 and 58 kWh/m², most likely 47" — with the assumption that drives the width named.
Prove it
Show one submission sheet where every performance number carries its source — surrogate or simulation — its inputs and its date, and where no surrogate figure stands alone.
How it works
A surrogate is a regression model trained on the outputs of a simulation; the surrogate-model card explains why it is fast and why it is silent outside its training range. Recent building-performance research shows what good practice looks like: an early-stage heating-and-cooling surrogate that reports an explicit validity domain and an uncertainty band with each prediction, and states that its fit scores describe interpolation within its own benchmark, not real buildings. The daylight surrogate on this card was trained on 2,880 simulations of a single-window shoebox room — useful for ranking options, and nothing like a courtyard. The method here is the studio's version of that discipline: the validity check is done by running the physics on the options that matter, rather than by the tool. It extends the climate-numbers card — screen fast, confirm properly — to the moment of handover. And it points at the checkpoint card: the simulation is where a human signs a performance claim.
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.