NOAA and Miami researchers derive a 129-day ceiling on weather forecasts from the energy budget

Toth says the identification at the core of the method came to the group intuitively: the time needed for incoming sunlight to replace all the energy in the atmosphere is the same quantity as the upper limit of predictability ^[12]. That identification is doing all the work: if it holds, the rest is arithmetic, and if it does not, the number does not stand either. He also says the technique involves no complex mathematics ^[13], which is both a virtue and a vulnerability, because nothing is concealed and nothing else is carrying the load.

The physical argument is about where information goes. Treat the atmosphere as a closed system whose initial state is perfectly known, and its deterministic dynamics would preserve that information indefinitely, permitting forecasts out to infinity ^[9]. The real atmosphere exchanges radiation, and the phases of photons in sunlight are entirely random and undeterminable, injecting quantum-scale uncertainty ^[10]. That noise erases retained information beginning at the smallest scales and working upward, and once the atmosphere's total energy has been completely replaced, predictive ability is gone ^[11]. The ceiling is therefore a clock on how long the initial condition survives being overwritten, rather than a statement about any model's error growth.

That gap between today's skill and the bound is what matters here. Subtract present-day skill from the bound and 115 days of headroom are unclaimed, putting the theoretical limit at roughly nine times current reach ^[1][2]. The stated uncertainty is narrow for a figure this far from anything demonstrated: between 122 and 136 days ^[3].

The result leaves open which technology, if any, reaches that limit. The authors say plainly that reaching the limit depends on future improvements in forecasting technologies ^[4], and the calculation is indifferent to whether those improvements are physics-based or learned. The only comparison in Physics World's coverage is a link to a companion report that physics-based models still beat AI at predicting extreme weather events ^[18]. The estimate is also global by construction, assembled from total atmospheric energy and the radiation fluxes at the atmosphere's upper boundary ^[3], so it cannot be spent on a single phenomenon. Tropical cyclones make the point: hurricane timing can now be predicted eight days ahead, which was deemed impossible a few decades ago ^[16].

If the equivalence holds, the argument about forecast horizons moves off physics and onto hardware and models, and the operative question becomes how much of those 115 days observation and computation can actually claim.