Engineering & Architecture Brief Published: 2026-11-15
Start by asking the vendor what the master algorithm does when one coupled solver misses its communication step. If the answer is that it waits, the whole twin is only as available as its slowest component. In that case one stalled solver freezes the plant view that operators were told to trust. The OpenUSD scene, the PINN surrogates and the Sim2Real pipeline all depend on how that one case is handled.
When a solver overruns its step Most industrial stacks couple their solvers through co-simulation interface standards such as FMI, and the standard says nothing about what your orchestrator does under deadline pressure. Ask the vendor to state, for a fixed-step real-time loop with a 10 ms macro step, whether a late slave blocks the master, gets extrapolated, or gets rolled back. Blocking keeps the physics correct but turns a single process hiccup into a system-wide stall. Extrapolation keeps the clock moving, but in stiff coupled systems it can add energy at the interface, so a thermal or hydraulic loop slowly drifts away from reality while every dashboard still shows green. Rollback is the clean answer only if every supplied model can save and restore its own state, and many third-party FMUs ship without that capability, so ask for the capability flags of each model in the bill of materials, not a general assurance.
Who holds state when the renderer dies Ask whether the OpenUSD stage is the system of record or only a view of it, because if live sensor state is written only into the Omniverse layer, a crashed Nucleus server or renderer erases what the twin knows about the plant, not just what it displays.
Surrogates that fail quietly Physics-informed neural networks make real-time multiphysics affordable, and that is why the failure mode most worth worrying about lives here: once the asset leaves the operating envelope the surrogate was trained on, it keeps returning smooth, plausible numbers that no longer track the physics. A crash is easy to catch. A confident surrogate running outside its training range is not. Ask what out-of-distribution detection runs alongside each surrogate, what threshold triggers a handover to the full-order solver, and whether that solver can actually keep up when called. By our estimate the full-order model runs one to two orders of magnitude slower than its surrogate, so a fallback that exists on paper may in practice push the loop past its deadline. The same question applies to Sim2Real robotics: a policy trained against a surrogate inherits its blind spots, and the vendor should be able to show which parts of the domain randomisation came from the full solver.
Spending the lab time you have With limited lab access, skip the fidelity demos the vendor has already rehearsed and use the time for fault injection instead. Kill one solver process, then the message broker, then the renderer, each mid-run. Here is the decision rule. If the remaining system does not degrade to a flagged, timestamped stale state within one macro step and recover without a full restart, rank that vendor below any competitor who passes, whatever their fidelity numbers. A vendor who cannot run this test on their own hardware has told you how their architecture behaves in production. Passing still leaves you a trade-off: graceful degradation means giving up some determinism, and someone on your side has to decide which subsystems may run stale and for how long, which is a plant operations decision, not one the vendor should make for you.
Technical FAQ: Does FMI compliance on the datasheet guarantee that models from different suppliers will coexist safely in one twin? No. Compliance only covers the interface. Whether a model supports state save and restore, variable step sizes or event handling is set by optional capability flags, and those flags decide whether rollback, recovery and deterministic replay will work at all. Ask for the model description files of every model you plan to couple and check those flags before you accept a vendor's word that the twin will survive a fault.