Phantom Non-Identifiability accepted at Sim2Science, NeurIPS 2026

Simplifying a model can make it invent problems that don't exist, and it'll look perfectly fine while doing it. That's the core of my paper "Phantom Non-Identifiability: When Simplifying a Coupled Simulator Fabricates an Observability Boundary", which has been accepted at the Sim2Science workshop at NeurIPS 2026 in Paris.

Scientists and engineers often simplify their simulation models to make them faster and then use those models to decide what to measure or how to control a process. I show that removing a single small term from a heat-and-moisture model makes it confidently claim that certain properties can't be measured, so you'd need to buy an extra sensor. In the real system, that limitation doesn't exist. The simplified model still fits the data well, so nothing raises an alarm, and the same blind spot carries over to neural network surrogates and controllers built on top of it.

The lesson: a model that fits well isn't automatically safe to make decisions with.

This work grew out of my time at the Institut für Textiltechnik (ITA) at RWTH Aachen University. Many thanks to Flemming Heuvel for the discussions that got me thinking about this problem.

I'll be presenting in Paris in December. If you work on simulation, inference, or scientific ML, let's talk.

Paper: OpenReview