The E3SM earth system model, like most climate models, uses a 300-second timestep for cloud microphysics. Rain forms, evaporates, accretes cloud water, breaks up — all resolved every five minutes. When the physics is stiff (large rates, fast transitions), the standard approach is brute-force substepping: divide the 300 seconds into smaller pieces and step through them one by one.
Dong et al. (arXiv:2603.11345) show that the 300-second timestep is fundamentally underresolved, and that brute-force substepping is the worst way to fix it. Higher-order Runge-Kutta methods with adaptive error control achieve the same accuracy more than 10x faster.
This is not a small improvement. It's a demonstration that the computational bottleneck in climate microphysics isn't hardware or parallelism — it's the mathematics of the time integrator. First-order forward Euler with substepping wastes computation by taking many small, inaccurate steps. A third- or fourth-order method takes fewer, larger, more accurate steps and arrives at the same answer with a fraction of the work.
The reason this hasn't been done before is institutional: climate models were developed when computational resources were scarce and code simplicity mattered more than mathematical sophistication. Forward Euler is three lines of code. Adaptive Runge-Kutta is a library. The engineering choice made in the 1990s became a performance ceiling in the 2020s.
The structural lesson: when a computation is slow, the bottleneck is often in the algorithm, not the hardware. More computation (smaller substeps, more processors) attacks the problem by scaling. Better mathematics (higher-order methods, adaptive control) attacks it by reducing the problem. Better mathematics beats more computation, and the factor is not 2x or 3x — it's 10x or more.