You can't damage a building to test your damage detection algorithm. But your algorithm needs damaged buildings to prove it works. The circularity is fundamental: structural health monitoring (SHM) methods require training data that responsible engineering won't produce.
A numerical benchmark (arXiv:2603.12069) builds the damaged building in simulation. A fixed-fixed steel beam under Euler-Bernoulli dynamics, with acceleration and displacement measurements that incorporate environmental and operational variations, measurement noise, and sensor faults. The damage types and severities are known exactly — something impossible in real structures, where you'd need to destructively verify the damage your algorithm claims to detect.
The tradeoff: the simulation is simpler than reality. A single-degree-of-freedom beam is not a skyscraper. Euler-Bernoulli theory ignores shear deformation. The environmental variations are parametric, not the chaotic weather and traffic loads a real structure endures. But the simplicity is the point. A benchmark isn't supposed to replicate reality. It's supposed to isolate variables. If your algorithm can't detect damage in a simplified beam with known physics, it certainly can't detect damage in a real structure with unknown everything.
The deeper insight: including sensor faults alongside structural damage forces algorithms to distinguish between “the structure changed” and “the measurement changed.” In real SHM, sensor drift, cable degradation, and power interruptions produce signals indistinguishable from actual damage. The benchmark doesn't just test detection. It tests discrimination.