The rain microphysics module in the Energy Exascale Earth System Model uses a 300-second time step -- a coarse resolution that would be unstable without ad hoc limiters clamping variables to physical ranges. Justin Dong and colleagues show that these limiters do their job too well: they prevent instability but simultaneously hide discretization errors. The simulation looks stable and produces plausible-looking output, but the underlying physics is underresolved. Reducing the time step to capture correct microphysics behavior without limiters increases computation time nearly 40-fold. Their solution -- higher-order Runge-Kutta integrators with adaptive time stepping -- achieves the required accuracy more than 10 times faster than the default scheme.
The structural problem is that stability and accuracy have been conflated. A limiter that prevents a negative rain mixing ratio is doing something physically reasonable -- rain cannot be negative -- but it is also absorbing the signal that would tell you your time step is too large. The simulation never crashes, never produces obviously wrong values, and therefore never triggers the alarm that would motivate a fix. This is a general failure mode in computational science: when the stabilization mechanism operates in the same space as the error, the error becomes invisible. The limiters are not wrong; they are doing exactly what they were designed to do. But their success at preventing catastrophe masks their failure at preventing inaccuracy. The adaptive integrators fix this by making accuracy, not stability, the criterion for step-size selection. The limiters become unnecessary because the time step is small enough to resolve the physics. Removing the safety net reveals the ground was always closer than it looked.
(arXiv:2603.11345)