arXiv subjects

Michael Denhard

Publications and source records attributed to Michael Denhard.

3 recordsLinked to original sources

arXiv2026

AICON: An operational global machine learning weather forecasting model

We introduce AICON, a global machine learning weather prediction (MLWP) model which generates forecasts at 13 km spatial resolution with a 3-hour time step, trained on the high-resolution, non-hydrostatic ICON-DREAM dataset. AICON is in full operational use at Deutscher Wetterdienst since 2nd of March 2026. The model employs a graph neural network (GNN) architecture with an encoder-processor-decoder structure, where node updates are performed using a graph attention mechanism. A key feature of AICON is its use of an icosahedral multi-mesh derived from the native grid of the ICON model, ensuring consistency with the training data. ICON's terrain-following vertical SLEVE coordinate is one of the major distinctions from existing emulators. AICON's training strategy prioritizes small-scale fidelity by avoiding autoregressive multi-step rollout and longer forecast horizons during training, a design choice motivated by the hypothesis that this approach preserves fine-scale features often damped in models optimized for longer-range forecasts. We describe the prognostic and diagnostic variables used for training, the transfer learning protocol employed to accelerate convergence, and the model's performance across a range of evaluation metrics. An extensive evaluation, including routine verification against observation, a tropical cyclone case and spectral analysis reveal the strengths and limitations in the representation of atmospheric variability across scales. Routine verification against observations demonstrates competitive skill relative to the operational ICON model, particularly for near-surface variables in the short to medium forecast range.

physics.ao-ph↗

arXiv2025

Arnoldi Singular Vector perturbations for machine learning weather prediction

Since weather forecasts are fundamentally uncertain, reliable decision making requires information on the likelihoods of future weather scenarios. We explore the sensitivity of machine learning weather prediction (MLWP) using the 24h Pangu Weather ML model of Huawei to errors in the initial conditions with a specific kind of Singular Vector (SV) perturbations. Our Arnoldi-SV (A-SV) method does not need linear nor adjoint model versions and is applicable to numerical weather prediction (NWP) as well as MLWP. It observes error growth within a given optimization time window by iteratively applying a forecast model to perturbed model states. This creates a Krylov subspace, implicitly based on a matrix operator, which approximates the local error growth. Each iteration adds new dimensions to the Krylov space and its leading right SVs are expected to turn into directions of growing errors. We show that A-SV indeed finds dynamically meaningful perturbation patterns for the 24h Pangu Weather model, which grow right from the beginning of the forecast rollout. These perturbations describe local unstable modes and could be a basis to initialize MLWP ensembles. Since we start A-SV from random noise perturbations, the algorithm transforms noise into perturbations conditioned on a given reference state - a process that is akin to the denoising process of the generic diffusion based ML model of GenCast, therefor we briefly discuss similarities and differences.

physics.ao-ph↗

arXiv2022

Krylov Methods for Adjoint-Free Singular Vector Based Perturbations in Dynamical Systems

The estimation of weather forecast uncertainty with ensemble systems requires a careful selection of perturbations to establish a reliable sampling of the error growth potential in the phase space of the model. Usually, the singular vectors of the tangent linear model propagator are used to identify the fastest growing modes (classical singular vector perturbation (SV) method). In this paper we present an efficient matrix-free block Krylov method for generating fast growing perturbations in high dimensional dynamical systems. A specific matrix containing the non-linear evolution of perturbations is introduced, which we call Evolved Increment Matrix (EIM). Instead of solving an equivalent eigenvalue problem, we use the Arnoldi method for a direct approximation of the leading singular vectors of this matrix, which however is never computed explicitly. This avoids linear and adjoint models but requires forecasts with the full non-linear system. The performance of the approximated perturbations is compared with singular vectors of a full EIM (not with the classical SV method). We show promising results for the Lorenz96 differential equations and a shallow water model, where we obtain good approximations of the fastest growing perturbations by using only a small number of Arnoldi iterations.

math.DS↗