Imagine you're navigating by sonar in a river, but the river's current changes direction while your ping is still traveling. If you assume the current was frozen the moment you sent the ping, your position estimate drifts. That's the core problem this paper tackles, except the "river" is Earth's atmosphere, the "ping" is an infrasound wave traveling hundreds or thousands of kilometers, and the "position estimate" is the state of a numerical weather model. The committed claim: for the first time, the authors compute full 4D (3D space + time) sensitivity kernels for infrasound observables — travel times, amplitudes — with respect to atmospheric wind and temperature fields that evolve during the wave's propagation. Previous infrasound simulation tools universally assumed a frozen atmosphere. This paper shows how to drop that assumption using adjoint methods, and demonstrates that the frozen-atmosphere shortcut introduces meaningful errors even when propagation spans just 2-3 atmospheric model time steps. The method belongs to the adjoint/variational family familiar from seismology and meteorological data assimilation. The authors formulate the forward acoustic propagation problem, derive its adjoint, and back-propagate sensitivity information through both space and time. The key insight is that the adjoint naturally handles time-varying coefficients at no extra architectural cost — it's the forward-model community that has been unnecessarily restricting itself to stationary atmospheres. Hardware demands are modest: the demonstration uses the Lorenz'96 toy model, not a full 3D atmospheric simulation. On the data assimilation side, the paper provides an important theoretical constraint: only variational methods (4D-Var and its ensemble hybrids) with assimilation windows at least as long as the infrasound propagation time can properly ingest these observations. Sequential methods like the ensemble Kalman filter with short windows cannot, because they'd need to freeze the observation operator within each window. This is a design constraint for operational weather centers considering infrasound data streams. The Lorenz'96 experiments are the integrity backbone. The authors assimilate synthetic circumnavigating Lamb wave travel times — the kind of signal generated by events like the 2022 Hunga Tonga eruption — into the toy model. They quantify the error introduced by both the linear approximation of the observation operator and the stationary-atmosphere approximation. The stationary approximation degrades analysis quality detectably even at 2-3 time steps of propagation, which for real atmospheres could mean just a few hours. The ladder comparison here is less about beating a competing method and more about demonstrating that an assumption the entire field makes is wrong and correctable. There's no named "SOTA" to beat; instead, the baseline is the community's standard practice of stationary-atmosphere infrasound modeling. The paper shows that practice is insufficient, provides the mathematical machinery to fix it, and quantifies the cost of not fixing it. What's missing is the obvious scale-up. The Lorenz'96 model has 40 variables. A real atmospheric model has millions of grid points. The adjoint infrastructure exists in principle (ECMWF's IFS has an adjoint), but coupling a full 3D acoustic propagation adjoint to an operational atmospheric adjoint is a serious engineering and computational challenge. The authors almost certainly know this is the next step — and almost certainly didn't attempt it because the compute and institutional coordination required are a multi-year, multi-team effort.
