Most plant simulations never get used on the shift

Every industrial software vendor now sells a twin. Some of them are even twins of something. LG Energy Solution’s Cheongju plant used a digital model of its 46-series cylindrical line to cut the time to bring new equipment to rate by more than half — a rare case where the software showed up in OEE rather than in a keynote. South Korea liked the result enough to talk about it as a national template. That is the ceiling. The floor is a Unity scene that has not been updated since the last kaizen.

Digital twins fail for boring reasons. The OT network is a museum. The PLC tags are unnamed. The process engineers who know why station 4 drifts on humid days are not in the digital-thread meeting. A twin that cannot ingest the real line’s data in a shift is a rendering. NVIDIA Omniverse and OpenUSD have made the rendering extraordinary. They have not made the instrumentation free.

One bottleneck, not a model of the whole site

The plants getting value are not “going twin.” They are modelling one bottleneck: a formation process, a paint shop, a changeover, a fleet of AMRs that keep colliding with a reality the layout drawing forgot. Foxconn’s Fii platform, Delta’s work in Isaac Sim, the TCS lights-out lab in Pune — the common pattern is a bounded cell with clean data, not a digital copy of 400,000 square metres.

If the twin cannot predict the scrap you had last Tuesday, do not ask it to run the plant on Thursday.

Agentic layers — software that does not just mirror the line but recommends or takes action — will make this more, not less, dependent on data hygiene. A wrong recommendation at rate is more expensive than no recommendation. The unfashionable work of naming tags, aligning clocks, and putting a process engineer in the loop is the actual AI programme.