Solar Inspection Stills from a Synthetic Twin

You are looking at a digital twin of a Jinko 72HC-style module, rendered in Isaac Sim 6 — not drone video, and not a fielded inspection product. Close-up RGB, same camera pose, defects you can read by eye.

What the service is

Sawatch builds the OpenUSD twin, randomizes the scene, renders RGB, and auto-labels classes so a vision model can train without a large labeled drone archive.

Digital twin

A procedural OpenUSD module twin: half-cut cells, silver busbars and fingers, center ribbon. Geometry and materials are owned in simulation, not scraped from flight logs.

Synthetic RGB

Isaac Sim 6 renders inspection-obvious stills. Lighting, soiling, cracks, and hotspots are authored in the twin so each class is visible in the image.

Auto-labeled export

Labels are generated with the image. Four classes on this demo: healthy, soiled, crack, hotspot. Ready for a training set — not a deployed detector.

Four classes, same pose

Each still is the same module framing. The only change is the defect class authored in the twin.

Healthy solar module with half-cut cells and silver busbars — **healthy** Half-cut cells, silver busbars and fingers, center ribbon. The clean baseline.

Solar module with dust soiling on visible busbars — **soiled** Same pose; dust blobs sitting on visible busbars.

Solar module with branching glass and cell fractures on the left — **crack** Branching glass/cell fractures on the left; grid still readable.

Solar module with a two-by-two amber and red hotspot cluster — **hotspot** Four amber/red cells in a 2×2 cluster.

Auto-labeled output

Hotspot still with YOLO-style class overlay and legend

YOLO-style box overlay on the hotspot still. The legend lists healthy, soiled, crack, and hotspot. This is class visualization on synthetic RGB — the auto-annotated view that ships with the image, not a trained production model.

Mapped back to the pipeline

The landing-page steps, grounded in this twin:

Ingest

Module geometry and PBR materials authored into OpenUSD.

Compose

Procedural twin in Isaac Sim; defect class and look randomized per frame.

Render

RGB stills at inspection scale from the GPU simulation.

Annotate

Class labels generated with the image for training export.

Build a twin like this

If you need labeled physical AI data from a digital twin — solar or otherwise — write us. This page is a demo of the service, not an operations product.

Contact