Surrogate Models & Fast Feedback

A surrogate model - also called a proxy model, metamodel, emulator or response-surface model - is a fast approximation of a slow, expensive process. In our world, that expensive process is usually a building performance simulation: an energy model, a daylight study, a CFD airflow run, a structural analysis. Each of these can take minutes to hours. A surrogate learns to mimic the relationship between the simulation's inputs and its outputs, so that instead of running the physics again, you ask the surrogate and get an answer in milliseconds.

The idea is not new or exotic - engineers have used response surfaces for decades - but modern machine learning makes surrogates far more capable and easy to build. The key mental picture: a surrogate is a function that has seen enough examples of 'these inputs produce this result' that it can guess the result for inputs it has not seen, as long as they are similar. It does not understand physics. It has learned a shape.

Why bother? Because speed unlocks a completely different way of working. When each answer costs an hour, you run a handful of studies and pick. When each answer costs a millisecond, you can explore thousands of variations, drive live sliders, and run automated optimisation. The surrogate does not just make the same work faster - it makes new kinds of work possible, which is exactly what the next lesson on optioneering builds on.

Building a surrogate has a clear recipe, and knowing it tells you exactly how much to trust the result. First, define the design space - the inputs that vary (say window ratio 20-60%, orientation 0-90 degrees, floor depth 8-16 m) and the output you care about (annual cooling energy). Second, sample it - choose many combinations of inputs, usually spread cleverly across the space so the examples cover it well. Third, run the real simulation for each sample. This is the slow part: maybe a few hundred honest runs. Fourth, train a model on that dataset of input-output pairs. Fifth, validate it - test its predictions against simulation runs it never saw, and measure the error.

That validation step is the one non-negotiable. A surrogate without a reported error is a rumour. A good workflow always asks: how close are the surrogate's answers to the real simulation on held-out cases, and is that close enough for the decision at hand?

A sketch of the pipeline in code-like pseudocode:

# 1-3: build the training data (the slow, one-time cost)
X, y = [], []
for design in sample_design_space(n=400):
    result = run_real_simulation(design)   # minutes each
    X.append(design.features)
    y.append(result.cooling_energy)

# 4: train the surrogate
surrogate = Model().fit(X, y)

# 5: validate on cases it never saw
error = evaluate(surrogate, X_holdout, y_holdout)

# now: near-instant predictions inside the design tool
prediction = surrogate.predict(new_design.features)  # milliseconds

Notice where the cost lives: entirely in steps 1-3, paid once. After that, predict is effectively free - which is the whole point.

You rarely meet a surrogate wearing a label that says 'surrogate' - you meet it as speed that should not be possible. When a tool updates an energy or daylight number the instant you drag a slider, when massing analysis in Autodesk Forma refreshes as you reshape a block, or when a Grasshopper definition scores thousands of variants in an optimisation run without waiting hours per variant, a surrogate is very often the reason. The physics did not suddenly get fast; a model learned to imitate it.

There are broadly three ways surrogates reach you. Baked into a tool, where the vendor trained it and you simply enjoy the responsiveness - convenient, but you must trust their training and rarely see the error band. Built in a parametric workflow, where you or a computational-design specialist trains a surrogate over a specific simulation for a specific project - more work, but you know exactly its range and accuracy. And as a growing set of plugins and libraries that automate the sampling-simulating-training pipeline, lowering the barrier between those two extremes.

The practical guidance shifts with each. For a baked-in surrogate, your job is to learn what the tool was calibrated for and stay inside it. For one you build, your job is honest validation - hold out simulation runs, measure the error, and report it. In both cases the surrogate earns its place only when its speed unlocks something you could not otherwise do: real-time steering, or a search over thousands of options. If a single simulation would answer your question just as well, run the simulation - a surrogate is worth building only when you will call it many, many times.

Once a surrogate exists, performance becomes interactive. This is where it stops being a technical curiosity and starts changing your day. In a parametric setup - Grasshopper or Dynamo with a surrogate plugged in - dragging a slider updates a predicted daylight or energy figure in real time, so you can literally feel which direction improves things. Tools like Autodesk Forma apply the same principle at the massing scale, updating sun, wind and energy estimates as you reshape a block.

Two behaviours become possible that were impractical before. The first is live design steering: you optimise by feel, watching numbers respond as you sketch, building intuition for the trade-offs. The second is automated search: because a surrogate can be evaluated thousands of times per second, an optimisation algorithm can sweep the whole design space and surface the best trade-offs for you - the engine behind AI optioneering (Lesson 7.4). Neither is feasible when every evaluation costs an hour of simulation time.

There is a subtler benefit too: fast feedback builds intuition. When you can see, live, that deepening a floorplate darkens its core or that turning a block eases its cooling load, you are not just picking a winner - you are learning the physics in your hands, the way a musician learns an instrument by playing it. Over dozens of projects that accumulated feel is arguably the surrogate's greatest gift, because it stays with you even when the tool is closed. The best designers who work this way do not become dependent on the numbers; they internalise them, and the live feedback quietly makes them better designers away from the screen.

The cultural shift matters as much as the technical one. Performance stops being a gate you pass at the end and becomes a material you shape from the start, the way you already shape form and light. That is the promise of fast feedback: not replacing your judgement, but feeding it continuously.

A surrogate is only trustworthy inside the range it was trained on. This is the single most important limit, and the one most often forgotten. Ask a surrogate about a design similar to its training examples and it interpolates - fills in a plausible answer between known points - usually well. Push it to inputs beyond what it saw - a much taller building, a climate it never trained on, a geometry outside the sampled range - and it extrapolates, which is a confident guess with little behind it. The number still appears, just as crisp as ever; nothing on screen warns you it is now unbacked. That silent confidence is the trap.

Other limits follow from how surrogates are built. They are frozen at training time: change the underlying simulation assumptions, the climate file, or the building typology and the old surrogate may no longer apply. They inherit any bias or gap in their training samples. And they smooth over sharp, threshold-like behaviour - a surrogate may miss a cliff-edge effect that the real physics captures, because it learned a smooth shape.

The defensive habits are straightforward. Know the surrogate's input ranges and stay inside them; treat any out-of-range prediction as unverified. Watch for reported error bands and respect them. And close the loop the same way as the last lesson: for high-stakes decisions, spot-check the surrogate against a fresh real simulation before you rely on it. A surrogate is a brilliant fast stand-in for the slow truth - as long as you remember it is a stand-in, and never mistake the imitation for the physics itself. The mark of someone who uses surrogates well is not that they trust the numbers more, but that they know precisely how far to trust each one, and where the edge of that trust lies.