Stand in the great room at four o'clock on an August afternoon and you can feel the design decision before you can name it, heat pooling against the west glass where the shades have hung drawn since June, which means the wall of windows the renderings celebrated now functions as a wall in the other sense: something you close off rather than look through. Somewhere in the plan set, a window schedule lists every unit by size and type with the confidence of engineering, yet most of those sizes were chosen the way they are always chosen, because the last house used them and the lumberyard stocks them and nobody ever had an afternoon to prove otherwise.
There is now a machine that has the afternoon, thousands of afternoons in fact, compressed into seconds and moving into the one phase of design where window decisions are still cheap to change.
Daylight simulation is not new. Architects have been able to model exactly how sun moves through a proposed building for decades, using physics-based engines like Radiance that trace light the way it actually behaves. What kept it out of single-family residential work was never accuracy but economics. A full annual simulation, the kind that scores a design on spatial daylight autonomy and annual sunlight exposure, the two metrics green certification actually grades, took hours per iteration and demanded a specialist to set up and interpret. Museums got daylight studies, while your 2,400-square-foot colonial did not.
A shortcut that learned physics
Machine-learning surrogate models changed the terms. Instead of simulating light physics fresh for every design tweak, researchers run a few thousand full simulations once, train a neural network on the results, and then let the network predict daylight performance for new designs in seconds. One Kuala Lumpur study needed 2,000 simulated cases to train a surrogate that predicts useful daylight illuminance alongside lighting, cooling, and solar-gain energy. A Stockholm project built its training set from real residential design scenarios and got a pix2pix image model predicting daylight factors fast enough to run live inside the CAD window, so compliance feedback arrives while the designer is still drawing. A 2026 peer-reviewed survey of the field reports models hitting R-squared values of 0.74 to 0.95 against held-out simulation data across daylight and glare metrics.
Read that last sentence carefully, because the honest version is narrower than the exciting version: those accuracy scores measure how well the network imitates the simulator, not how well the simulator imitates your house, and the difference matters more than the marketing admits.
Software vendors noticed, and Autodesk's Forma now runs machine-learning daylight analysis in real time during massing studies, scoring facade surfaces against an overcast-sky model and flagging the stretches of glass that will never see decent light. Its newer Forma Building Design layer, built around what Autodesk calls a neural CAD foundation model trained on 3D building data, promises generative layout tools that evaluate options against daylight as they draw. Some of that is shipping, some of it remains roadmap, and per industry reporting the generative pieces arrive later this year or in 2026, so believe the shipped parts.
Meanwhile cove.tool took an adjacent path: reduced-order energy modeling that the Department of Energy describes as delivering energy estimates in practically real time across entire parametric sweeps, deliberately sidestepping the BIM-to-energy-model geometry translation failures that used to eat consultants' weeks. What matters for small residential practices is the price tag. Press coverage put a five-person team under $3,500 a year, which is the number that moved performance analysis from hired-gun territory into something a twelve-person firm runs in-house on a Tuesday.
Running the payback math
Here is where I part ways with the brochure: I ran the payback arithmetic, and you deserve to watch it fail.
Methodology first. Federal data sets the baseline, since the Energy Information Administration's residential survey found the average American home spends 1,105 kilowatt-hours a year on lighting, roughly a tenth of household electricity. Assume, generously, that daylight-optimized design cuts lighting energy by 30 percent. That is an illustrative assumption, not a measured outcome, and I am labeling it as such because the literature range is wide and house-specific. Do the arithmetic: 1,105 times 0.30 is about 330 kilowatt-hours a year per home. That is the prize the energy-savings pitch is fighting over, and it shrinks every year as LEDs take over. EIA's 2024 survey data, released this March, shows LEDs have become the dominant indoor bulb type in American homes, which means the lighting load that daylighting displaces is a melting ice cube.
So no, the software does not pay for itself in kilowatt-hours on any single house, and any vendor implying otherwise is selling you a feeling. Look elsewhere for the honest case, because it is stronger than the energy math and it lives in the rooms you inhabit rather than the meter you read, in comfort and glare and the afternoon sun you never have to fight. It is the west-facing great room that never gets built. It is the glare study that catches, in schematic design when glass is still lines on a screen, the afternoon blast that would otherwise be discovered the first August after move-in and fixed with motorized shades, window film, or re-glazing at a cost nobody wants to say out loud. It is a small firm walking into a client meeting with physics instead of taste, which wins work, and one deleted glass wall repays the subscription, because the kilowatt-hours were never the point.
There is a second honest number the industry should sit with. cove.tool's founders have claimed their optimization cuts initial construction cost 2 to 3 percent or buys 40 percent better performance for a 3 percent premium, per a 2018 industry profile. I found no independent audit of those figures anywhere. Treat them as the company's story about itself until someone checks.
Arguing the other side at full strength
A surrogate model is a compression of its training data, not physics, and most published daylight surrogates were trained on offices, classrooms, or generic apartment blocks. Your house is none of those. It has a neighbor's two-story looming twelve feet off the property line, interior paint the model never saw, and occupants who will operate the blinds according to moods no dataset captures. Blinds behavior alone can swing realized daylight performance enormously, and no surrogate trained on empty simulated rooms knows how you live.
Then there is the subtler hazard, which worries me more than any accuracy table. When the simulation took hours and needed a specialist, its slowness enforced humility. Everyone in the room understood it was an approximation wearing a lab coat. When the prediction is instant and rendered as a beautiful heatmap inside the drawing tool, it invites a different relationship, and the designer stops questioning it, while the tool democratizes analysis and simultaneously democratizes misplaced certainty that a profession in love with convincing renderings may not notice.
Finally, the metrics themselves can be gamed without trying. Spatial daylight autonomy rewards bringing light deep into the plan. Annual sunlight exposure penalizes too much direct sun. Read separately, a designer can maximize one while ignoring the other and produce a bright, glaring room the metrics technically bless. They must be read together, and nothing in the software forces that discipline, even though the physics got faster and judgment did not.
What this analysis could not prove
Limits, stated plainly. I found no verified case study of a specific single-family home where a surrogate daylight model changed what got built; the residential evidence base is datasets and shipping tools, not occupied-home measurements. That $3,500-a-year pricing figure is 2018 press and unconfirmed today. The 30 percent lighting-savings assumption in my payback math is illustrative. EIA's detailed lighting figure is 2015 vintage, with 2024 end-use detail not due until spring 2027. Surrogate accuracy scores come from the papers' own test sets, not independent replication. And Autodesk's generative layout features remain roadmap, not product, for most users.
What to do with this
If you are building custom or remodeling with an architect, ask one pointed question during schematic design: did you simulate daylight, or eyeball it? Then ask what changed because of it. A good answer names a moved window, a resized overhang, a killed skylight. "We have a feel for this orientation" is now a choice, not a necessity, and you are allowed to price it accordingly.
If you run a small residential practice, the economic shift is real even where the energy math is thin. Consultant-grade daylight analysis used to mean a five-figure engagement per project, and subscription tools now put it in-house for roughly the cost of one avoided glazing mistake, because the breakeven was never the power bill. It is the client meeting where you show physics.
If you are buying a production home, none of this helps you directly, and I will not pretend otherwise, because the simulation belongs to the design phase and your house already survived value engineering three years ago. You cannot re-simulate that plan. But you can visit at four in the afternoon in August. Stand in the great room. Feel where the heat pools. No AI can fix a west-facing glass wall either. Orientation is still destiny, and some truths remain analog.