Searcharxiv⌕ Search

arXiv subjects

Hongqin Zhang

Publications and source records attributed to Hongqin Zhang.

4 recordsLinked to original sources

One2N: Recursive refinement breaks the resolution barrier for Earth-system AI

Earth-system AI models approach kilometre-scale resolution, but halving grid spacing quadruples the grid-point count and end-to-end training cost. We introduce One2N, a recursive incremental framework extending a coarse model through $N$ levels of cross-resolution residual correction. At each level the parent is frozen and excluded from backpropagation, so its activations and optimizer states are never retained; only a tier-specific module is trained. A single refinement level on the NeuralGCM-0.7$^\circ$ baseline yields the two-level One2N-0.07$^\circ$, trained on one 80 GB GPU. Against HRES-0.07$^\circ$ and ERA5-0.25$^\circ$ analyses it matches its parent's tropospheric skill while reducing upper-stratospheric errors and resolving finer structure. Tropical-cyclone track errors nearly match the parent's and are lower than Pangu-Weather-0.25$^\circ$ at medium range; RMSE is lower than that model's on 93.46% of common targets. One2N thus turns an existing forecasting system, or an inexpensively trained coarse baseline, into a reusable foundation for high-resolution development, breaking the resolution barrier.

physics.ao-ph↗

Regularized Sequential Reconstructions of Dynamic Nonsmooth Sources from Sparse Boundary Flux Data

We consider an inverse source problem for the heat equation on the unit disk $Ω\subset\mathbb R^2$ with flux measurements at finitely many boundary points. The source is piecewise constant in time on prescribed intervals and, on each interval, is the characteristic function of an unknown region with Lipschitz boundary. The measurements on each interval split into the contribution of the current source and a history term from earlier sources; subtracting the latter leaves a linear inverse problem for a forward operator $F_τ$ that is densely defined and unbounded on $L^2(Ω)$. We prove that $F_τ$ is not closable, so the classical regularization theory for closed unbounded operators does not apply. We regularize instead with a Hilbert-scale penalty that compactifies it. The resulting Galerkin--Tikhonov reconstruction is sequential: each source is recovered from the flux observed up to the end of its interval and from the previously reconstructed sources. We prove its convergence and uniqueness, show that earlier reconstruction errors propagate into later data with exponential decay in time, and establish stability estimates with respect to the number of sensors. A numerical experiment on noisy measurements shows the feasibility of the sequential approximations, refined by Fourier approximation and boundary correction by a neural network.

math.NA↗

Handling errors in four-dimensional variational data assimilation by balancing the degrees of freedom and the model constraints: A new approach

For many years, strongly and weakly constrained approaches were the only options to deal with errors in four-dimensional variational data assimilation (4DVar), with the aim of balancing the degrees of freedom and model constraints. Strong model constraints were imposed to reduce the degrees of freedom encountered when optimizing the strongly constrained 4DVar problem, and it was assumed that the models were perfect. The weakly constrained approach sought to distinguish initial errors from model errors, and to correct them separately using weak model constraints. Our proposed i4DVar* method exploits the hidden mechanism that corrects initial and model errors simultaneously in the strongly constrained 4DVar. The i4DVar* method divides the assimilation window into several sub-windows, each of which has a unique integral and flow-dependent correction term to simultaneously handle the initial and model errors over a relatively short period. To overcome the high degrees of freedom of the weakly constrained 4DVar, for the first time we use ensemble simulations not only to solve the 4DVar optimization problem, but also to formulate this method. Thus, the i4DVar* problem is solvable even if there are many degrees of freedom. We experimentally show that i4DVar* provides superior performance with much lower computational costs than existing methods, and is simple to implement.

physics.ao-ph↗

A New Approach for 4DVar Data Assimilation

Four-dimensional variational data assimilation (4DVar) has become an increasingly important tool in data science with wide applications in many engineering and scientific fields such as geoscience1-12, biology13 and the financial industry14. The 4DVar seeks a solution that minimizes the departure from the background field and the mismatch between the forecast trajectory and the observations within an assimilation window. The current state-of-the-art 4DVar offers only two choices by using different forms of the forecast model: the strong- and weak-constrained 4DVar approaches15-16. The former ignores the model error and only corrects the initial condition error at the expense of reduced accuracy; while the latter accounts for both the initial and model errors and corrects them separately, which increases computational costs and uncertainty. To overcome these limitations, here we develop an integral correcting 4DVar (i4DVar) approach by treating all errors as a whole and correcting them simultaneously and indiscriminately. To achieve that, a novel exponentially decaying function is proposed to characterize the error evolution and correct it at each time step in the i4DVar. As a result, the i4DVar greatly enhances the capability of the strong-constrained 4DVar for correcting the model error while also overcomes the limitation of the weak-constrained 4DVar for being prohibitively expensive with added uncertainty. Numerical experiments with the Lorenz model show that the i4DVar significantly outperforms the existing 4DVar approaches. It has the potential to be applied in many scientific and engineering fields and industrial sectors in the big data era because of its ease of implementation and superior performance.

physics.data-an↗