AI for Performance Prediction

The cost of a design change rises steeply as a project matures. In the first sketches, moving a building's orientation, adjusting its depth, or trimming the glazing on a west face costs nothing but a rethink. Once you are in construction documents, the same change ripples through structure, services and coordination. This is the famous MacLeamy curve idea: the ability to influence performance is highest exactly when our knowledge of performance is usually lowest.

Traditional building performance simulation - a full energy model in EnergyPlus, an annual daylight study with a raytracer, a finite-element structural analysis - is accurate but slow and specialised. It often needs a detailed model and hours of setup and run time, so in practice it happens late, once per milestone, run by a specialist. AI-based prediction attacks the timing problem. Instead of a slow physics run, a trained model looks at your massing and inputs and returns an estimate almost instantly. Suddenly you can ask 'what if the atrium were narrower?' and see an answer before the thought cools.

The headline benefit is not that AI is more accurate than simulation - it usually is not - but that it is fast enough to be in the room while you design. It moves performance from a verdict at the end to a conversation throughout.

There is no magic here, and understanding the mechanism is what lets you judge the output. A performance-prediction model is trained on examples: many past designs (or synthetic ones generated on purpose) each paired with the result of a real, trusted simulation. The model learns the statistical relationship between design inputs - orientation, window-to-wall ratio, floor depth, envelope U-values, shading, climate - and the outcome, such as annual heating and cooling energy or a daylight autonomy score.

Once trained, the model no longer runs physics. It pattern-matches: given a new design that resembles what it has seen, it predicts the likely result. This is why the same idea appears under several names you will meet - a surrogate model, a proxy model, a metamodel, or a data-driven predictor (the next lesson, Surrogate Models & Fast Feedback, goes deep on exactly this). The important consequence is simple: the model is only as good as the data it learned from, and it is trustworthy mainly for designs that resemble that data.

Some tools bake this in so you barely see it. Autodesk Forma, for example, gives near-real-time analysis of sunlight, wind, noise and operational energy for early massing - fast approximations meant to guide, not to certify. Others are research tools or plugins where you can inspect the model more directly. Either way, the mental model to hold is: learned correlation, delivered instantly - not a fresh physical calculation each time.

It helps to know where this shows up in real practice, because the label 'AI' is often invisible. The clearest example is Autodesk Forma (the successor to Spacemaker), which gives near-real-time feedback on sunlight hours, daylight potential, wind, noise and operational energy as you sketch massing on a site - fast approximations designed to guide early decisions, explicitly not to certify them. Energy platforms such as cove.tool, Sefaira and Autodesk Insight sit a step later, offering quick whole-building energy and cost estimates that increasingly lean on fast approximation to feel responsive.

On the daylight side, tools built on the Ladybug Tools ecosystem and engines like Radiance remain the accurate ground truth, while lighter estimators give instant read-outs during design. For structure, ML-based estimators are more experimental - used in research and some early-feasibility tools to guess member sizes or feasibility from geometry - and here the accuracy caveat bites hardest, because structure is life-safety.

Two honest points about the 2026 state of these tools. First, maturity varies wildly: massing-scale energy and daylight prediction is genuinely useful today, while some 'AI performance' claims are thin wrappers around ordinary calculation, and a few are frankly overstated. Ask what a tool was calibrated against and how its accuracy was tested; a vendor who cannot answer is selling confidence, not capability. Second, the interface hides the mechanism - a slick real-time number gives no hint whether it came from a validated engine or a rough proxy. Part of using these tools well is knowing, for each one, which it is, and calibrating your trust to match. The skill is not chasing the newest badge but understanding what sits behind the number on your screen.

Here is the honest heart of the lesson. A predicted number is an estimate produced by pattern-matching, and it can be confidently wrong in ways that are hard to spot. It can be biased if the training set under-represented your building type or climate. It can be badly off for an unusual geometry it never saw. And it always carries an error band that the tidy single number on screen quietly hides. Treating a prediction as if it were a validated simulation result is the classic, dangerous mistake.

The discipline is to let the stakes set the scrutiny - the same principle that runs through this whole course. When you are comparing options - which of six massing schemes daylights best - the ranking is often reliable even if the absolute kWh figures are not, and that is enough to steer you. When the number will be relied on - sizing HVAC plant, proving a design meets an energy code or a daylight standard, signing off structural adequacy - you must validate with a proper simulation or a qualified engineer. Fast prediction is superb for exploration and direction; it is not a substitute for verification and proof.

A practical way to say it to yourself:

Use the prediction to CHOOSE a direction.
Use a real simulation to CONFIRM a decision.
Never let the estimate be the evidence for a claim you must defend.

There is a professional dimension here that is easy to underrate. The numbers a designer puts their name to carry responsibility - to the client who budgets around them, to the code official who relies on them, and to whoever ultimately occupies the building. An AI prediction has no such accountability; it cannot be liable, cannot be questioned in a meeting, and cannot stand behind its own figure. When you present a number, you are standing behind it, whatever produced it. That alone is reason enough to keep the line between predicted and validated crisp in your own records and your reports: not as bureaucratic caution, but because the difference is exactly the difference between a number you can defend and one you merely hope is right.

Hold that line and AI prediction is one of the most useful things in your kit. Blur it and you are shipping unverified numbers with a professional's name on them.

In practice, AI prediction lives at the front of the process and hands off to real simulation later. A typical rhythm looks like this. During concept and massing, you use an instant-feedback tool to sweep options - rotating the block, testing depths, trying shading strategies - watching predicted energy and daylight move in real time. You are hunting for the shape of the right answer, and approximate numbers are perfect for that.

As a scheme firms up, the role of AI shifts from steering to flagging. A prediction that a west facade will overheat, or that a floorplate is too deep to daylight, is a cue to investigate - a smoke alarm, not a fire report. You then commission a proper study on the decisions that survived. The prediction did its job by making sure the expensive simulation is spent on a design already pointed in a sensible direction.

A few habits keep this honest. Always know what the tool was trained or calibrated for, and distrust it outside that range. Sanity-check predictions against intuition and rules of thumb - if a number violates physics you understand, the model is wrong, not physics. Record which figures are predicted and which are simulated in your own notes, so a rough estimate never silently graduates into a reported result. Done this way, prediction compresses weeks of trial-and-error into an afternoon, while the numbers you actually stake your reputation on stay properly earned. That combination - explore fast with AI, prove slowly with simulation - is the whole art of this module.