Where AI belongs in solar design, and where it doesn’t

Every solar design tool says it uses AI now. That tells an engineering team very little. What matters is which tasks the AI is doing, and what happens when it gets one wrong.

How does Solesca decide where to use AI?

At Solesca we automate nearly every step between an address and an interconnection plan set. Before we automate a task, we decide whether it should run on an AI model or on deterministic math.

This post explains how we make that call, and why it matters if you're comparing design tools or thinking about building your own.

What do design teams say about AI?

We talk to a lot of design teams, and the feedback on AI is consistent. Engineers like it when they can check its work. They stop trusting it when they can't.

One designer described the obstruction detection in the tool his team used before as all or nothing. It either selected nothing, or it invented obstructions that weren't on the roof. Another watched a detection run on a cluttered roof and said he'd much rather check the AI's work than draw every little box by hand.

We also hear that obstructions are where the hours go. One C&I team told us keep-outs take more than half of their design time. So that's the first place to put AI to work. It's also a place where a wrong answer is easy to spot and quick to fix.

Where does AI earn its place in solar design?

AI models are good at reading messy inputs, like a satellite image or a PDF. They work on probability, so we use them where a person can review the result in seconds.

Rooftop obstruction detection. ARGUS, our computer vision model, finds the vents, skylights and HVAC units on a roof in under 30 seconds. It labels each type and applies the setback and height from your obstruction template. A sensitivity slider lets you decide how sure the model has to be before it draws anything.

Spec sheet parsing. Drop in a PDF cut sheet and Solesca builds the custom module or inverter file for you, ready to use in a design. No more typing values out of a datasheet.

In both cases the AI does the tedious first pass and the designer makes the final call.

Which design work has to give the same answer every time?

Anything that gets stamped, submitted to a utility, or built. In Solesca those outputs run on deterministic engineering math, with no AI model involved.

Terrain tracking. Trackers and fixed tilt tables sit on bare-earth elevation, and the layout solves for pile reveals and slope limits. The energy model then simulates that geometry in 3D.

Trees and wetlands. Canopy height and wetlands come in as map layers from published datasets. Tree and wetland obstructions carry heights and setbacks that you set.

AC and DC stringing. Strings and home runs are routed automatically. You get a schedule with lengths, voltage drop and wire sizes for every string.

Engineering mode. Conductors and gear are sized against the NEC, and each row of the schedule cites the code reason for its size. A design that doesn't comply can't enter engineering mode. Engineering mode covers US projects under the NEC today.

We show the code reason on every row because engineers told us they won't sign off on a black box. The same inputs give you the same plan set today and on revision six.

Why do in-house AI design tools stall?

If you have good engineers and access to a large language model, building your own design assistant looks tempting. A first demo comes together fast. The trouble starts after the demo.

A generative model can give two different answers to the same question. That's fine for a first pass at a roof scan. It's a problem for a fuse size. To make an in-house tool safe, your team has to work out task by task where a model can be trusted, then build a deterministic engine for everything else. After that, someone has to keep it all current as code cycles, utility requirements and equipment change. That someone is usually your best engineer.

We've spent years on that sorting work. Our COO interconnected hundreds of megawatts of projects. We do this so your engineers can spend their time on projects.

See it on your own site

Send us an address or a project you're working on right now. We'll run it live so you can check the obstruction detection, the stringing and the plan set against your own work.