Appendix (Supplementary Material) SOON: Symmetric Orthogonal Operator Network for Global Abstract Accurate global Subseasonal-to-Seasonal (S2S) climate forecasting is critical for disaster preparedness and resource management, yet it remains challenging due to chaotic atmospheric dynamics. Existing models predominantly treat atmospheric fields as isotropic images, conflating the distinct physical processes of zonal wave propagation and meridional transport, and leading to suboptimal modeling of anisotropic dynamics. In this paper, we propose the Symmetric Orthogonal Operator Network (SOON) for global S2S climate forecasting. It couples: (1) an Anisotropic Embedding strategy that tokenizes the global grid into latitudinal rings, preserving the integrity of zonal periodic structures; and (2) a stack of SOON Blocks that models the alternating interaction of Zonal and Meridional Operators via a symmetric decomposition, structurally mitigating discretization errors inherent in long-term integration. Extensive experiments on the Earth Reanalysis 5 dataset demonstrate that SOON establishes a new state-of-the-art, significantly outperforming existing methods in both forecasting accuracy and computational efficiency. marginparsep has been altered. topmargin has been altered. marginparpush has been altered. The page layout violates the ICML style. Please do not change the page layout, or include packages like geometry, savetrees, or fullpage, which change it for you. We’re not able to reliably undo arbitrary changes to the style. Please remove the offending package(s), or layout-changing commands and try again. SOON: Symmetric Orthogonal Operator Network for Global Subseasonal-to-Seasonal Climate Forecasting Ziyu Zhou 1 Tian Zhou 2 Shiyu Wang James Kwok 3 Yuxuan Liang 1 † † footnotetext: 1 The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China 2 DAMO Academy, Alibaba Group, Hangzhou, China 3 The Hong Kong University of Science and Technology, Hong Kong SAR, China. Correspondence to: Yuxuan Liang . Preprint. August 10, 2026. 1 Introduction Subseasonal-to-Seasonal (S2S) climate forecasting, predicting atmospheric conditions from 2 to 6 weeks in advance, is vital for decision-making in energy scheduling, agriculture, and disaster management ( 43 ; 45 ; 19 ) . However, forecasting on this timescale remains a grand challenge as it occupies the gap between initial-condition-dependent weather prediction and boundary-condition-dependent climate projection, where neither signal alone provides sufficient predictability ( 22 ) . Although data-driven models have advanced medium-range (up to 2 weeks) forecasting ( 3 ; 15 ; 32 ) , their direct extensions to S2S protocols typically suffer from rapid error accumulation ( 26 ; 21 ) . This performance gap highlights a fundamental mismatch between the isotropic inductive biases of standard architectures and the anisotropic nature of S2S dynamics. Figure 1: The anisotropic nature of global atmospheric dynamics. Global atmospheric circulation is governed by the coupling of two orthogonal physical processes: Zonal Dynamics, dominated by periodic wave propagation (e.g., planetary Rossby waves), and Meridional Transport, driven by the north-south exchange of heat and momentum (e.g., eddy heat fluxes). Recently, data-driven models have achieved strong performance in global S2S climate forecasting ( 28 ) . Efforts have adapted to the spherical grid structure via geometry-inspired circular patching strategies ( 19 ) or employed equivariant message passing on spherical meshes to handle vector field symmetries ( 20 ) . Additionally, spectral approaches have been explored to integrate teleconnection modeling with spherical harmonic representations ( 21 ) and incorporate Fourier-based embeddings for enhancing wave pattern representations ( 17 ) . Collectively, these works have propelled S2S forecasting from empirical statistical models toward end-to-end neural modeling, thereby introducing essential inductive biases for spherical geometry and global dependencies ( 6 ; 48 ) . A fundamental limitation of existing S2S methods is their reliance on isotropic image modeling, which treats global weather fields as generic 2D images where latitudinal and longitudinal dimensions are processed uniformly. From a physical perspective, as illustrated in Figure 1 , global atmospheric circulation is naturally decomposed into two coupled but distinct mechanisms: zonal periodic wave propagation ( 34 ; 14 ) and meridional transport processes ( 31 ; 40 ) . By treating these dimensions uniformly, standard baselines implicitly conflate these distinct physical roles, resulting in insufficient modeling of the true atmospheric dynamics. Crucially, this isotropic assumption also neglects the severe projection distortion inherent in equirectangular grids, causing standard architectures to struggle with the rapid phase variations of planetary waves, particularly in high-latitude regions. Specifically, existing S2S forecasting models are limited by isotropic modeling assumptions, and so standard 2D patching ( 17 ) , circular patching ( 19 ) , and spherical harmonic aggregation ( 21 ) fail to explicitly decouple propagation from transport, weakening the representation of complex wave-flow interactions and leading to suboptimal forecasting performance. Instead of relying on isotropic modeling, we propose the S ymmetric O rthogonal O perator N etwork (SOON) , which is designed based on the principles of physical process decoupling and symmetric operator decomposition for effective global S2S climate forecasting. SOON first employs an Anisotropic Embedding strategy that tokenizes the grid into latitudinal rings, preserving the integrity of zonal periodic structures while structurally distinguishing them from meridional representations. The core SOON Blocks explicitly model the alternating interaction of these processes via a symmetric decomposition, with a Zonal Operator capturing global spectral wave dynamics and a Meridional Operator modeling local spatial transport. The operators are designed to induce a geometrically orthogonal subspace structure in the representation, thereby aligning the model more naturally with the anisotropic nature of global atmospheric dynamics. Crucially, each SOON block introduces structural symmetry to mitigate the accumulation of discretization errors over the long S2S forecasting horizon. We further adopt a phase-stable normalization scheme tailored for extended forecast windows. Finally, the encoded representation is passed through a lightweight MLP-based Decoder that enables forecasts on all atmospheric variables, after which the entire model is trained end-to-end. To our knowledge, SOON is the first S2S forecasting framework that explicitly formulates zonal–meridional decoupling as a symmetric operator-splitting problem in the latent space. Our main contributions are as follows: • We identify that prevalent S2S forecasting models fundamentally violate the anisotropic nature of atmospheric dynamics through isotropic modeling, providing a unified perspective on their limitations regarding latitude-longitude decoupling and operator decomposition. • We propose SOON for S2S climate forecasting, which integrates Anisotropic Embedding with a symmetric composition of Zonal and Meridional Operators that is theoretically and structurally aligned with long-term atmospheric dynamics at the S2S timescale. • Extensive experiments on the Earth Reanalysis 5 (ERA5) dataset demonstrate that SOON establishes a new state-of-the-art, outperforming existing methods in both forecasting accuracy and computational efficiency. Figure 2: The main architecture of our proposed SOON. The Anisotropic Embedding projects atmospheric initial states into latitudinal ring tokens by compressing the longitudinal dimension into feature channels. These tokens are then evolved through a backbone of several stacked SOON Blocks. Specifically, each block sequentially applies two weight-shared Zonal Operators and a Meridional Operator to explicitly decouple wave propagation and transport dynamics. Finally, a Decoder maps the evolved features to bi-weekly forecast fields. 2 Preliminaries 2.1 Problem Definition We consider global bi-weekly forecasting of K meteorological variables defined on a latitude–longitude grid. Let \Omega=[-90^{\circ},90^{\circ}]\times[-180^{\circ},180^{\circ}] be the spherical coordinate domain, discretized into H latitude bands and W longitude points. We represent the spatial coordinates by \bm{\mathcal{P}}\in\mathbb{R}^{H\times W\times 2} with \bm{\mathcal{P}}_{h,w,:}=(\varphi_{h},\lambda_{w})\in\Omega , where \varphi_{h} and \lambda_{w} denote latitude and longitude, respectively. At time index t (in days), the global atmospheric state is a tensor \bm{X}_{t}\in\mathbb{R}^{K\times H\times W} . Following common S2S benchmarking protocols ( 25 ; 26 ) , given an initial condition (\bm{\mathcal{P}},\bm{X}_{t_{0}}) , our goal is to learn a predictor \mathcal{F}_{\Theta} that directly outputs the two verification targets for weeks 3–4 and weeks 5–6: (\hat{\bm{Y}}_{15:28},\,\hat{\bm{Y}}_{29:42})=\mathcal{F}_{\Theta}(\bm{\mathcal{P}},\bm{X}_{t_{0}}), (1) where \hat{\bm{Y}}_{15:28}\in\mathbb{R}^{K\times H\times W} and \hat{\bm{Y}}_{29:42}\in\mathbb{R}^{K\times H\times W} denote the predicted averages over days 15 – 28 (weeks 3–4) and days 29 – 42 (weeks 5–6) relative to the initialization day t_{0} , respectively. Our setting focuses on directly estimating these two evaluation-aligned bi-weekly targets from the initialization, in alignment with prior S2S forecasting works ( 21 ; 19 ; 20 ; 6 ; 25 ) . A summary of the existing work on weather forecasting and S2S climate forecasting can be found in Appendix A . 2.2 Operator Splitting for Non-commutative Dynamics Atmospheric evolution can be abstracted as a dynamical system defined by the differential equation \frac{\partial\mathbf{u}}{\partial t}\equiv\mathcal{F}(\mathbf{u})=\mathcal{A}(\mathbf{u})+\mathcal{B}(\mathbf{u}) , where \mathbf{u}(t) is the global state vector, \mathcal{A} and \mathcal{B} represent different physical processes (e.g., zonal wave dynamics and meridional transport) ( 34 ; 14 ; 31 ) . For easier exposition, we use the linear notation \mathcal{F}(\mathbf{u})=(\mathcal{A}+\mathcal{B})\mathbf{u} , in which the exact time- \tau flow can be written as \mathbf{u}(t+\tau)=\exp(\tau(\mathcal{A}+\mathcal{B}))\,\mathbf{u}(t) . When the sub-operators do not commute ( [\mathcal{A},\mathcal{B}]\neq 0 ), this flow cannot be factorized into separate evolutions, motivating operator splitting methods ( 39 ; 37 ; 23 ) . Operator splitting approximates the coupled evolution by alternating simpler sub-evolutions governed by \mathcal{A} and \mathcal{B} . Different splitting orders correspond to different discretizations of the same underlying dynamics and lead to distinct error behaviors. In long-horizon forecasting, these error properties matter because local approximation errors can accumulate under repeated composition. In the following, we introduce two representative operator splitting methods. Lie–Trotter splitting ( 39 ) . The first-order decomposition applies the two subflows sequentially, i.e., \mathcal{S}_{\text{Lie}}(\tau)=\exp(\tau\mathcal{B})\exp(\tau\mathcal{A}) . By the Baker–Campbell–Hausdorff (BCH) formula ( 2 ) , its local truncation error is \mathcal{O}(\tau^{2}) and is controlled by the commutator [\mathcal{A},\mathcal{B}] ( 4 ) . Strang splitting ( 37 ) . A symmetric second-order decomposition enforces time-reversal symmetry via \mathcal{S}_{\mathrm{Strang}}(\tau)=\exp\!\left(\tfrac{\tau}{2}\mathcal{A}\right)\exp(\tau\mathcal{B})\exp\!\left(\tfrac{\tau}{2}\mathcal{A}\right), (2) which cancels the leading second-order BCH terms and yields an \mathcal{O}(\tau^{3}) local error (for a sufficiently small time step \tau ) involving nested commutators ( 23 ) . In practice, symmetric decomposition is often preferred when modeling non-commutative systems, as it typically reduces systematic bias and slows down error accumulation compared with sequential decomposition ( 11 ; 4 ) . 3 Methodology Figure 2 illustrates the overall architecture of SOON, which is designed to align neural operations with the anisotropic nature of atmospheric dynamics. The framework is composed of three strategic modules: (i) Anisotropic Embedding , which projects the global grid into a sequence of latitudinal ring tokens by compressing longitudinal spatial information into feature channels; (ii) A stack of Symmetric Operator Blocks (SOON Blocks), which function as high-order neural solvers via a symmetric Strang splitting scheme, explicitly employing a Zonal Operator to capture global spectral wave dynamics and a Meridional Operator to model local spatial transport; (iii) A lightweight Decoder , which projects the evolved latent representations back to the physical space to generate the final bi-weekly forecasts. The entire architecture is trained end-to-end to minimize the prediction error over the S2S timescale. 3.1 Anisotropic Embedding Recent data-driven S2S models adopt various patch-based embedding strategies to tokenize global gridded fields, such as circular patching ( 19 ) , spherical harmonic embedding ( 21 ) , and Fourier embedding ( 17 ) . However, as in conventional approaches ( 30 ; 6 ) , they still treat the atmosphere as an isotropic image, which can disrupt continuous zonal wave structures that are essential for planetary-scale dynamics. This limitation motivates us to design a tokenization strategy that preserves zonal periodicity while decoupling it from the meridional transport processes. To this end, we introduce Anisotropic Embedding. Instead of extracting local patches, we project the entire latitudinal ring at each latitude index h into a single high-dimensional latent token. Formally, given an input state \bm{X}_{t_{0}}\in\mathbb{R}^{K\times H\times W} , we apply a 2D convolution, operating over the spatial latitude ( H ) and longitude ( W ) dimensions, with a kernel size (1,W) to compress the longitudinal dimension, followed by flattening and normalization: \bm{E}^{(0)}=\mathrm{RMSN}\left(\mathrm{Flat}\left(\mathrm{Conv}_{1\times W}(\bm{X}_{t_{0}})\right)\right)\in\mathbb{R}^{H\times C}, (3) where \mathrm{Conv}_{1\times W} denotes the convolution operation mapping K input variables to C fixed feature channels. \mathrm{Flat}(\cdot) represents the flattening operation that reshapes the tensor spatial dimension into a sequence of length H , and \mathrm{RMSN}(\cdot) denotes Root Mean Square Normalization (RMSNorm) ( 47 ) , which stabilizes the latent features without recentering. Consequently, the sequence dimension H of \bm{E}^{(0)} strictly corresponds to the physical meridional domain, while the feature dimension C encodes the latent zonal spectral information. Unlike local patching that fragments the grid ( 30 ; 17 ) , this global aggregation treats the full longitudinal span as a unified entity, thereby implicitly retaining the intrinsic cyclic topology of atmospheric waves within the latent feature space. 3.2 SOON Block After Anisotropic Embedding, the latent representation \bm{E}^{(0)} is processed through a backbone of L stacked SOON blocks. Each block functions as a neural operator solver with alternating spectral and spatial operations that advance the atmospheric state by a discrete time step \tau . We design the internal structure of the block to mimic the symmetric Strang splitting scheme defined in Eq. ( 2 ) ( \mathcal{S}_{\mathrm{Strang}}(\tau)=e^{\frac{\tau}{2}\mathcal{A}}e^{\tau\mathcal{B}}e^{\frac{\tau}{2}\mathcal{A}} ). This aims to minimize the discretization error accumulated over long forecast horizons due to non-commutativity of atmospheric operators ( 23 ; 11 ) . Formally, let \bm{E}^{(l-1)}\in\mathbb{R}^{H\times C} be the input to the l -th block. We define two operators: (i) Zonal Operator \mathcal{Z}(\cdot) , and (ii) Meridional Operator \mathcal{M}(\cdot) , that correspond to \mathcal{A} and \mathcal{B} in Eq. ( 2 ), respectively. Details of these operators will be introduced later in this section. The block updates the representation via a symmetric composition path followed by a Feedforward Network ( \mathrm{FFN(\cdot)} ) with residual connection: \displaystyle\bm{E}^{\prime}\! \displaystyle=\!\bm{E}^{(l-1)}\!+\!\bm{\gamma}_{1}\!\odot\!\left(\mathcal{Z}\!\circ\!\mathcal{M}\!\circ\!\mathcal{Z}\right)\!\left(\!\mathrm{RMSN}(\bm{E}^{(l-1)})\!\right), (4) \displaystyle\bm{E}^{(l)}\! \displaystyle=\!\bm{E}^{\prime}\!+\!\bm{\gamma}_{2}\!\odot\!\mathrm{FFN}(\mathrm{RMSN}(\bm{E}^{\prime})), (5) where \circ denotes operator composition and \odot is the element-wise Hadamard product. The parameters \bm{\gamma}_{1},\bm{\gamma}_{2}\in\mathbb{R}^{C} correspond to learnable LayerScale ( 38 ) vectors that adaptively modulate the signal amplitude in residual branches to facilitate stable training. In Eq. ( 4 ), the Zonal Operator \mathcal{Z} is applied twice with weight sharing. With this structural symmetry constraint, the SOON block satisfies time-reversal symmetry ( 11 ; 4 ) , which automatically eliminates all even-order error terms (e.g., \mathcal{O}(\tau^{2}) ) in the splitting approximation, thereby directly addressing the error accumulation challenge in long-term forecasting. The following Theorem provides a theoretical guarantee on error control. Theorem 3.1 ( Local Truncation Error of SOON Block ) . Let \mathcal{L}_{\mathcal{Z}} and \mathcal{L}_{\mathcal{M}} be the infinitesimal generators corresponding to the Zonal and Meridional operators, respectively. Let \mathcal{S}_{\mathrm{SOON}}(\tau) be the evolution operator of one SOON block approximating the exact dynamics e^{\tau(\mathcal{L}_{\mathcal{Z}}+\mathcal{L}_{\mathcal{M}})} . By enforcing the symmetric composition \mathcal{Z}\circ\mathcal{M}\circ\mathcal{Z} , the local truncation error is of third order, i.e., \|\mathcal{S}_{\mathrm{SOON}}(\tau)-e^{\tau(\mathcal{L}_{\mathcal{Z}}+\mathcal{L}_{\mathcal{M}})}\|=\mathcal{O}(\tau^{3}) , which is strictly superior to the \mathcal{O}(\tau^{2}) error of standard sequential stacking based on Lie-Trotter splitting. Proof. Based on the symmetric Baker-Campbell-Hausdorff (BCH) ( 2 ) formula, the symmetric structure strictly cancels the second-order commutator error terms ( [\mathcal{L},\mathcal{K}] ) inherent in sequential splitting. See Appendix C.1 for the complete derivation. ∎ While Layer Normalization (LayerNorm) ( 1 ) has been widely adopted in prior S2S forecasting architectures ( 19 ; 20 ; 21 ) , the following Proposition identifies that its mean-centering operation physically disrupts energy conservation of atmospheric waves. Proposition 3.2 . Let \mathbf{z}\in\mathbb{R}^{C} be the latent representation of one of the latitudinal ring tokens in \bm{E}^{(0)} from Eq. ( 3 ), and let \hat{\mathbf{z}}=\mathcal{F}(\mathbf{z}) be its discrete Fourier spectrum, where \mathcal{F} denotes the Discrete Fourier Transform (DFT). LayerNorm introduces spectral distortion by eliminating the zero-frequency component (background energy), i.e., |\mathcal{F}(\text{LayerNorm}(\mathbf{z}))_{0}|\equiv 0 , where \mathcal{F}(\cdot)_{0} represents the 0 -th frequency component of the DFT. In contrast, the following Proposition shows that RMSNorm acts as a uniform spectral scaler. This is critical for S2S forecasting because phase information encodes the spatial positioning and structural topology of planetary waves, which must be preserved to prevent trajectory divergence over long horizons. Accordingly, SOON consistently uses RMSNorm instead of LayerNorm to stabilize training. Proposition 3.3 ( Spectral Fidelity of RMSNorm ) . RMSNorm preserves the phase angles \arg(\mathcal{F}(\text{RMSN}(\mathbf{z}))_{k})=\arg(\hat{\mathbf{z}}_{k}) for all wave numbers k . Proof. See Appendix C.2 and Appendix C.3 for detailed proof of the propositions above based on the Parseval’s theorem ( 29 ) and Fourier Transform. ∎ In the following, we detail the implementations of the Zonal and Meridional Operators in each SOON block. Zonal Operator \mathcal{Z} . It is designed to model global wave propagation (e.g., Rossby waves) along the longitudinal direction. Since our anisotropic embedding compresses longitudinal information into the C feature channels, zonal dynamics are naturally captured by mixing information across the channel dimension. We implement this via a gating mechanism in the frequency domain: \mathcal{Z}(\bm{N})=\mathrm{iFFT}\left(\bm{W}_{\mathcal{Z}}\odot\mathrm{FFT}(\bm{N})\right), (6) where \bm{N}\in\mathbb{R}^{H\times C} is the normalized input \mathrm{RMSN}(\bm{E}^{(l-1)}) in Eq. ( 4 ), \mathrm{FFT}(\cdot) and \mathrm{iFFT}(\cdot) denote the one-dimensional Fast Fourier Transform and its inverse applied along the feature dimension, respectively, and \bm{W}_{\mathcal{Z}}\in\mathbb{C}^{C} is a learnable complex weight initialized near identity. By operating in the frequency domain, this operator captures global periodic dependencies with an effective global receptive field ( 49 ) , preserving phase information critical for wave dynamics while maintaining O(C\log C) time complexity. Meridional Operator \mathcal{M} . It targets the latitude dimension H , where atmospheric dynamics are dominated by local transport (i.e., advection) and thermodynamic mixing rather than global periodicity. We model this via a spatial Gated Linear Unit (GLU) convolution ( 8 ) : \mathcal{M}(\bm{N})=\left(\mathrm{Conv}(\bm{N})\odot\sigma(\mathrm{Conv}_{\mathrm{gate}}(\bm{N}))\right)\bm{W}_{\mathrm{proj}}, (7) where \sigma(\cdot) is the sigmoid function, \mathrm{Conv}(\cdot) and \mathrm{Conv}_{\mathrm{gate}}(\cdot) are two distinct 1D convolution layers, both utilizing a fixed kernel size k sliding over the meridional sequence dimension H , and \bm{W}_{\mathrm{proj}} is a learnable parameter. The element-wise product \odot serves as a gating mechanism, allowing the model to dynamically control the information flux, akin to the nonlinear advection terms in fluid dynamic equations ( 40 ) , facilitating the learning of local meridional transport patterns. 3.3 Decoder After passing through L SOON blocks, the latent representation \bm{E}^{(L)}\in\mathbb{R}^{H\times C} contains the evolved atmospheric state in the compressed anisotropic space. The Decoder projects this representation back to the original physical grid to generate the final forecasts. We employ a linear projection layer followed by a tensor reshaping operation to reconstruct the longitudinal dimension W from the C feature channels: (\hat{\bm{Y}}_{15:28},\,\hat{\bm{Y}}_{29:42})=\mathrm{Reshape}\left(\mathrm{Linear}(\bm{E}^{(L)})\right), (8) where the linear layer projects the feature dimension from C to 2\times K\times W . Finally, we obtain two tensors, \hat{\bm{Y}}_{15:28} and \hat{\bm{Y}}_{29:42} , each with dimension \mathbb{R}^{K\times H\times W} , representing the two target bi-weekly mean fields corresponding to weeks 3–4 and weeks 5–6, respectively. Training Objective. The model is trained in an end-to-end manner, minimizing the discrepancy between the predicted bi-weekly means and the ground truth. Based on the target definition in Eq. ( 1 ), we define the overall objective \mathcal{L} as the average squared error over both forecast windows: \mathcal{L}\!=\!\frac{1}{2KHW}\!\left(\left\|\hat{\bm{Y}}_{15:28}\!-\!\bm{Y}_{15:28}\right\|_{\mathcal{L}_{2}}^{2}\!\!\!+\!\!\left\|\hat{\bm{Y}}_{29:42}\!-\!\bm{Y}_{29:42}\right\|_{\mathcal{L}_{2}}^{2}\right)\!, (9) where \bm{Y}_{15:28} and \bm{Y}_{29:42} denote the ground truth averaged fields for weeks 3–4 and weeks 5–6, respectively. To account for the spherical geometry of the Earth, the norm \|\cdot\|_{\mathcal{L}_{2}}^{2} represents the latitude-weighted Mean Squared Error, where the error at each latitude h is weighted by w_{h}=\cos(\varphi_{h}) following ( 19 ) . 3.4 Computational Complexity Analysis Standard Transformer-based models ( 9 ; 27 ) incur a quadratic computational cost of \mathcal{O}((HW)^{2}C) , scaling quadratically with the longitudinal resolution ( W^{2} ). In contrast, SOON achieves a complexity of \mathcal{O}(H(C^{2}+C\log C+Ck)) (detailed derivations in Appendix B ). Crucially, this complexity is independent of the longitudinal width W due to our anisotropic embedding. This strictly linear scaling with latitude H provides a significant efficiency advantage for high-resolution global forecasting compared to isotropic architectures. Table 1: Global S2S climate forecasting results. The best result is in bold , and the second-best is underlined . The lower RMSE and higher ACC indicate better performance. We report several key pressure-level variables ( z500 , z850 , t500 , and t850 ) and single-level variables ( t2m , u10 , and v10 ). N/A indicates that a baseline does not predict the corresponding variable(s). Model RMSE ( \downarrow ) ACC ( \uparrow ) Pressure Level Single Level Pressure Level Single Level z500(gpm) z850(gpm) t500(K) t850(K) t2m(K) u10(m/s) v10(m/s) z500(gpm) z850(gpm) t500(K) t850(K) t2m(K) u10(m/s) v10(m/s) Weeks 3-4 (Days 15-28) CMA 69.833 47.056 2.598 2.918 4.221 N/A N/A 0.968 0.919 0.975 0.974 0.960 N/A N/A NCEP 68.441 44.787 2.225 2.742 4.028 N/A N/A 0.969 0.921 0.979 0.977 0.966 N/A N/A ECMWF 63.674 42.349 2.067 2.391 3.232 N/A N/A 0.972 0.929 0.982 0.983 0.971 N/A N/A [1.2pt/2.5pt] Pangu 66.224 37.959 2.271 2.569 N/A 2.431 1.984 0.963 0.926 0.966 0.963 N/A 0.812 0.686 FCN2 62.755 41.020 2.093 2.390 N/A 2.328 1.896 0.947 0.896 0.966 0.957 N/A 0.830 0.712 [1.2pt/2.5pt] Transformer 59.748 38.119 2.038 2.531 3.662 2.244 1.793 0.978 0.948 0.983 0.981 0.978 0.840 0.738 FNO 68.308 41.014 2.302 2.896 3.912 3.442 2.328 0.965 0.942 0.960 0.942 0.972 0.837 0.742 ViT 62.122 39.474 2.249 2.852 4.378 2.379 1.926 0.976 0.945 0.979 0.976 0.969 0.814 0.678 [1.2pt/2.5pt] ClimaX 59.064 38.420 2.142 3.242 5.342 2.360 2.025 0.978 0.945 0.981 0.971 0.956 0.807 0.649 ClimODE 68.767 39.326 2.518 3.147 3.782 2.326 1.793 0.970 0.946 0.972 0.970 0.975 0.827 0.738 CirT 55.505 34.345 1.931 2.265 5.478 2.253 1.873 0.980 0.955 0.984 0.983 0.958 0.835 0.716 EIMP 66.498 41.820 2.312 2.800 4.785 3.683 2.137 0.981 0.956 0.983 0.984 0.974 0.672 0.654 TelePiT 59.772 35.199 2.035 2.448 4.002 2.122 1.746 0.977 0.953 0.982 0.980 0.974 0.854 0.741 SOON 51.956 33.453 1.824 2.145 3.255 1.994 1.621 0.982 0.958 0.985 0.985 0.982 0.872 0.783 Weeks 5-6 (Days 29-42) CMA 71.400 48.583 2.639 2.933 5.521 N/A N/A 0.966 0.912 0.974 0.974 0.950 N/A N/A NCEP 71.332 46.834 2.343 2.833 4.406 N/A N/A 0.967 0.913 0.977 0.976 0.952 N/A N/A ECMWF 68.116 45.330 2.177 2.511 4.080 N/A N/A 0.968 0.916 0.979 0.981 0.970 N/A N/A [1.2pt/2.5pt] Pangu 76.939 47.041 2.829 2.998 N/A 2.679 2.104 0.956 0.911 0.958 0.957 N/A 0.783 0.655 FCN2 66.531 43.469 2.250 2.567 N/A 2.479 1.980 0.943 0.889 0.963 0.953 N/A 0.812 0.691 [1.2pt/2.5pt] Transformer 59.970 38.433 2.065 2.568 3.454 2.279 1.824 0.977 0.947 0.983 0.981 0.980 0.836 0.725 FNO 72.021 44.327 2.564 3.334 4.301 2.426 2.084 0.960 0.949 0.971 0.971 0.968 0.840 0.720 ViT 63.031 39.825 2.287 2.974 4.469 2.397 1.979 0.976 0.944 0.979 0.970 0.968 0.815 0.641 [1.2pt/2.5pt] ClimaX 58.712 38.271 2.141 3.300 5.354 2.394 2.031 0.978 0.945 0.982 0.971 0.956 0.800 0.643 ClimODE 68.218 39.287 2.554 3.213 4.094 2.336 1.844 0.971 0.946 0.972 0.969 0.971 0.829 0.717 CirT 55.214 33.940 1.920 2.265 5.652 2.251 1.826 0.980 0.956 0.984 0.983 0.954 0.835 0.726 EIMP 66.856 41.926 2.329 2.858 5.018 2.555 2.155 0.981 0.957 0.983 0.983 0.972 0.792 0.651 TelePiT 58.770 35.164 1.953 2.451 4.056 2.120 1.750 0.978 0.953 0.982 0.981 0.975 0.854 0.741 SOON 52.268 33.437 1.830 2.156 3.231 1.993 1.606 0.982 0.958 0.985 0.985 0.982 0.873 0.786 4 Experiments 4.1 Experimental Setup Dataset. Consistent with the established benchmarking protocols ( 19 ; 20 ; 21 ) , we evaluate SOON on the ERA5 reanalysis dataset ( 13 ) , which is widely used as a reference for global atmospheric fields. We regrid the data to a 1.5^{\circ} resolution, resulting in a 121\times 240 latitude–longitude grid. Following ( 19 ; 20 ) , we use 63 variables in total. This includes six pressure-level variables (geopotential z , specific humidity q , temperature t , u-wind u , v-wind v , and vertical velocity w ) at ten pressure levels (10, 50, 100, 200, 300, 500, 700, 850, 925, and 1000 hPa), as well as three surface variables (2-meter temperature t2m , 10-meter u-wind u10 , and 10-meter v-wind v10 ). Following ( 19 ; 20 ) , we split the dataset chronologically: using 1979 to 2016 for training, 2017 for validation, and 2018 for testing. Detailed descriptions of the dataset can be found in Appendix D.1 . Baselines. We benchmark SOON against a wide range of state-of-the-art methods, which can be divided into two categories: (1) Operational NWP Systems : including the physics-based models from CMA ( 46 ) , NCEP ( 35 ) , and ECMWF ( 24 ) . (2) Data-driven Models : encompassing foundation models such as PanguWeather (Pangu) ( 3 ) , FourCastNetV2 (FCN2) ( 30 ) , ClimaX ( 27 ) , and specialized S2S models including CirT ( 19 ) , EIMP ( 20 ) , TelePiT ( 21 ) , and ClimODE ( 42 ) . Additionally, we compare against general-purpose backbone models: Transformer ( 41 ) , Fourier Neural Operator (FNO) ( 18 ) , and Vision Transformer (ViT) ( 9 ) . We adapt these general-purpose models to the S2S forecasting task by modifying their input and output projection heads to accept the global grid tensor while maintaining their original backbone architectures to process the tokenized spatiotemporal sequences. We do not compare with TianQuan-S2S ( 17 ) , which uses a different forecasting horizon, and OmniCast ( 28 ) , which operates at a different spatial resolution than ours. See details in Appendix D.2 . Metrics. Following existing evaluation standards ( 33 ; 26 ) , we employ the Latitude-weighted Root Mean Square Error (RMSE) and Anomaly Correlation Coefficient (ACC) to quantify performance ( 26 ) , defined as: \begin{aligned} \text{RMSE}&=\sqrt{\frac{1}{HW}\sum_{h,w}\alpha(h)\,(\hat{\bm{X}}_{h,w}-\bm{X}_{h,w})^{2}},\\ \text{ACC}&=\frac{\sum_{h,w}\alpha(h)\,\hat{\bm{X}}^{\prime}_{h,w}\,\bm{X}^{\prime}_{h,w}}{\sqrt{\sum_{h,w}\alpha(h)\,(\hat{\bm{X}}^{\prime}_{h,w})^{2}}\;\sqrt{\sum_{h,w}\alpha(h)\,(\bm{X}^{\prime}_{h,w})^{2}}},\end{aligned} (10) where \alpha(h)=\cos(\phi_{h})\big/\left(\frac{1}{H}\sum_{h^{\prime}}\cos(\phi_{h^{\prime}})\right) represents the latitude weighting factor. \bm{X} and \hat{\bm{X}} denote the ground truth and prediction for a given bi-weekly verification window. The anomalies are computed as \bm{X}^{\prime}=\bm{X}-\bm{C} and \hat{\bm{X}}^{\prime}=\hat{\bm{X}}-\bm{C} , where \bm{C} is the observational climatology. Table 2: Ablation study quantifying the contribution of each SOON component. We evaluate variants by removing Anisotropic Embedding (w/o AE), Zonal Operator (w/o ZO), or Meridional Operator (w/o MO), and by replacing RMSNorm (w/o RN) with LayerNorm (w/ LN). Model RMSE ( \downarrow ) ACC ( \uparrow ) Pressure Level Single Level Pressure Level Single Level z500(gpm) z850(gpm) t500(K) t850(K) t2m(K) u10(m/s) v10(m/s) z500(gpm) z850(gpm) t500(K) t850(K) t2m(K) u10(m/s) v10(m/s) Weeks 3-4 SOON 51.956 33.453 1.824 2.145 3.255 1.994 1.621 0.982 0.958 0.985 0.985 0.982 0.872 0.783 w/o AE 63.472 41.836 2.618 3.041 4.356 3.604 3.052 0.958 0.921 0.946 0.914 0.952 0.828 0.732 w/o ZO 61.207 39.118 2.467 2.742 4.061 3.053 3.441 0.970 0.939 0.953 0.934 0.961 0.818 0.721 w/o MO 64.218 42.507 2.334 3.176 4.119 3.284 3.121 0.955 0.926 0.962 0.908 0.958 0.844 0.764 w/ unshared ZO 54.124 35.021 1.940 2.217 3.421 2.128 1.822 0.968 0.956 0.977 0.980 0.976 0.860 0.776 w/o RN, w/ LN 62.948 40.917 2.412 2.957 4.047 3.067 3.302 0.968 0.932 0.959 0.924 0.966 0.856 0.777 Weeks 5-6 SOON 52.268 33.437 1.830 2.156 3.231 1.993 1.606 0.982 0.958 0.985 0.985 0.982 0.873 0.786 w/o AE 63.761 43.284 2.512 2.913 3.982 3.617 2.621 0.965 0.938 0.954 0.942 0.964 0.834 0.742 w/o ZO 61.543 40.643 2.721 2.783 3.889 3.318 2.952 0.974 0.936 0.951 0.953 0.963 0.821 0.728 w/o MO 62.872 44.019 2.407 2.792 3.771 3.544 2.386 0.961 0.949 0.960 0.938 0.971 0.846 0.759 w/ unshared ZO 55.213 36.221 1.960 2.322 3.501 2.012 1.708 0.970 0.950 0.978 0.980 0.980 0.862 0.766 w/o RN, w/ LN 63.318 42.674 2.581 2.979 4.107 3.702 2.561 0.970 0.952 0.967 0.936 0.973 0.857 0.772 Figure 3: The global RMSE distribution of t850 with lead times weeks 3-4 in the testing set. Implementation Details. For PanguWeather and FourCastNetV2, we utilize the official pre-trained weights provided by the ECMWF project through API (see Appendix D.2 for details). For the other baselines (except the Operational NWP Systems CMA, NCEP, ECMWF), we train them from scratch using a unified experimental setting to ensure fair comparison (batch size 32, hidden dimension 256, 16 attention heads, and trained for 20 epochs using the AdamW optimizer with an initial learning rate of 0.01). The implementation is based on PyTorch Lightning and training is conducted on 8 NVIDIA H20 80G GPUs. Inference is performed on a single NVIDIA H100 80G GPU. More implementation details are in Appendix D.3 . 4.2 Overall Performance Table 1 shows the quantitative comparison against operational NWP systems and state-of-the-art data-driven baselines. As can be seen, SOON consistently achieves the best performance across almost all evaluated metrics, surpassing the leading operational systems, pretrained climate foundation models, general purpose models, as well as recent data-driven climate forecasting models by a substantial margin. Notably, SOON demonstrates exceptional robustness in the extended weeks 5–6 window, maintaining high ACC and low RMSE while other methods suffer from significant degradation. This consistent superiority validates that our anisotropic modeling of wave-flow dynamics combined with the symmetric operator splitting scheme can effectively mitigate the accumulation of errors inherent in long-term iterative forecasting. Furthermore, SOON exhibits comprehensive strength across diverse physical regimes, ranking first in both stable upper-air fields and highly chaotic surface variables. We also evaluate SOON with various values of L (the number of SOON blocks), and the results are presented in Appendix E.1 . Additional comparisons on spectral energy distribution are provided in Appendix E.2 . Figure 4: RMSE variation across latitudinal bands. SOON demonstrates superior accuracy, achieving the lowest errors near the poles. 4.3 Ablation Study To validate the effectiveness of our design choices, Table 2 compares SOON with variants with several crucial components removed. The detailed experimental settings are in Appendix D.3 . The results show that removing Anisotropic Embedding causes the most severe performance degradation across all variables, confirming that standard isotropic patching disrupts continuous zonal wave structures essential for S2S dynamics. Similarly, omitting either Zonal Operator or Meridional Operator leads to significant error increases, validating the necessity of explicitly decoupling spectral wave propagation from spatial transport. Additionally, using an unshared-weight Zonal Operator also leads to worse performance, indicating that the symmetric structure of SOON numerically reduces error. Furthermore, the variant using LayerNorm consistently underperforms the RMSNorm baseline, which empirically supports our theoretical analysis that RMSNorm better preserves background energy for long-term integration. Figure 5: RMSE comparison between SOON and data-driven models on variable z and t across different pressure levels. 4.4 Empirical Analysis Global Visualization. Figure 3 visualizes the spatial distribution of RMSE for t850 . A striking pattern emerges among state-of-the-art data-driven baselines: they all exhibit significant error accumulation in high-latitude regions, as highlighted by the red boxes. We attribute this common failure to their reliance on isotropic modeling assumptions. This approach inherently struggles with the severe projection distortion near the poles. However, SOON maintains a low error profile globally. By adopting anisotropic embedding, SOON respects the distinct physical topology of latitudinal rings, immunizing predictions against high-latitude distortion. See more results in Appendix E.3 . Figure 6: Efficiency analysis on the ERA5 dataset. Metrics are measured with batch size 1 and hidden dimension 256 for a fair comparison. Lower values indicate higher efficiency. Quantitative Latitudinal Analysis. Figure 4 breaks down the RMSE across four latitudinal bands ranging from the equator ( 0^{\circ} ) to the pole ( 90^{\circ} ). A clear divergence in performance trends is observed: standard isotropic models exhibit a sharp upward error trend as latitude increases, confirming their vulnerability to grid convergence and projection distortion. In stark contrast, SOON demonstrates a unique inverse trend, where forecasting error often decreases or remains minimal in the polar region. This quantitative evidence reinforces that our anisotropic ring embedding not only resolves the singularity issue but effectively leverages the strong zonal continuity of polar dynamics, turning geometric constraints into a structural advantage. Performance Variation across Pressure Levels. Figure 5 visualizes the normalized RMSE for variables ( z ) and ( t ) across 9 pressure levels. SOON consistently exhibits superior accuracy from the stratosphere to the near-surface. In contrast, baselines show distinct performance degradation, particularly in the upper atmosphere (50–200 hPa). This vertical consistency confirms that our anisotropic embedding and symmetric operator design effectively preserve the physical coherence of atmospheric profiles across all altitudes. See more visualizations in Appendix E.4 . Additionally, we present daily forecasting performance within the bi-weekly horizon in Appendix E.5 , and monthly forecasting performance over the full testing set in Appendix E.6 . Efficiency Comparison. Figure 6 highlights the efficiency of SOON. Our model substantially reduces computational burdens compared to existing baselines. Notably, SOON achieves an inference speedup of over 2 \times and 5 \times compared to CirT and ClimaX, respectively. This efficiency stems from our anisotropic design, replacing quadratic attention mechanisms with linear-complexity Zonal and Meridional operators, ensuring scalability for practical forecasting. 5 Conclusion and Future Work This paper addresses the fundamental mismatch between existing isotropic baselines and anisotropic atmospheric dynamics in S2S forecasting. We propose SOON, a physics-aligned architecture that explicitly decouples zonal waves from meridional transport via anisotropic ring embedding and a symmetric Strang splitting scheme. Empirical results on ERA5 confirm that SOON establishes a new state-of-the-art in both accuracy and efficiency. We discuss the limitations of SOON in Appendix F . Future work will focus on integrating generative frameworks to quantify forecast uncertainty essential for chaotic S2S timescales. Impact Statements This work contributes to global subseasonal-to-seasonal forecasting. We do not foresee any negative ethical consequences or societal impacts from it. Table of Contents Appendix A Related Works A.1 Data-driven Weather Forecasting Spatio-temporal data, together with time series data, are prevalent across various domains of daily life, among which weather data are closely related to human activities ( 10 ; 44 ; 50 ) . Recently, deep learning has emerged as a computationally efficient alternative to traditional Numerical Weather Prediction (NWP) systems ( 43 ; 45 ) . While early works focused on regional tasks ( 36 ; 12 ) , the release of ERA5 ( 13 ) has shifted focus to global forecasting. Foundational benchmarks ( 33 ) paved the way for models that rival operational NWP systems (e.g., IFS) in the medium range. Prominent architectures include Adaptive Fourier Neural Operators (FourCastNet ( 30 ) ), Graph Neural Networks (GraphCast ( 15 ) , AIFS ( 16 ) ), and Transformers (PanguWeather ( 3 ) , FengWu ( 5 ) , FuXi ( 7 ) ). Recent research also explores probabilistic forecasting via diffusion models ( 32 ) . However, the performance of these medium-range models often degrades at subseasonal scales due to the lack of physical constraints and error accumulation. A.2 Data-driven S2S Forecasting Subseasonal-to-Seasonal (S2S) forecasting (2–6 weeks) remains a grand challenge due to chaotic atmospheric dynamics. While traditional dynamical models (e.g., ECMWF ( 24 ) ) face computational bottlenecks, recent global data-driven models have shown promise. ClimaX ( 27 ) adapts a vision foundation model for climate tasks via fine-tuning. FuXi-S2S ( 6 ) employs a cascade architecture to forecast daily averaged statistics. CirT ( 19 ) introduces circular patching strategies to better handle spherical boundary conditions. EIMP ( 20 ) leverages graph neural networks with rotation-equivariant message passing to model vector fields. TelePiT ( 21 ) combines spherical harmonics with physics-informed neural ODEs to capture multi-scale teleconnections, while TianQuan-S2S ( 17 ) further enhances predictability by incorporating climatology priors and noise injection. However, these approaches predominantly rely on isotropic modeling assumptions that neglect the distinct physical nature of zonal waves versus meridional transport, often suffering from high computational costs and error accumulation. In contrast, our proposed SOON introduces a physics-aligned anisotropic architecture that explicitly disentangles these dynamics, achieving superior computational efficiency and forecasting stability. Appendix B Theoretical Model Complexity Here, we provide the step-by-step derivation of the time complexity for both standard Transformers and SOON. For standard Transformer-based models (e.g., ClimaX ( 27 ) , ViT ( 9 ) , CirT ( 19 ) ), the input grid H\times W is typically tokenized into T\propto HW patches. Consequently, the computational cost is dominated by the global self-attention mechanism, which scales as \mathcal{O}(T^{2}C) , and feed-forward networks scaling as \mathcal{O}(TC^{2}) . Substituting T , the total complexity becomes \mathcal{O}((HW)^{2}C+HWC^{2})\approx\mathcal{O}(H^{2}W^{2}C) , exhibiting a quadratic dependency on both latitudinal ( H ) and longitudinal ( W ) resolutions. In contrast, SOON utilizes anisotropic embedding to compress the longitudinal dimension, resulting in a reduced sequence length of H . The computational burden of a single SOON block comprises three specific components: the Zonal Operator, which performs FFT and iFFT along the feature dimension with complexity \mathcal{O}(HC\log C) ; the Meridional Operator, which applies depthwise 1D convolution with kernel size k along the sequence length, costing \mathcal{O}(HCk) ; and the point-wise feed-forward networks, contributing \mathcal{O}(HC^{2}) . Summing these terms yields a total complexity of \mathcal{O}(H(C^{2}+C\log C+Ck)) . This derivation explicitly demonstrates that the longitudinal width W does not factor into the backbone’s computational cost, confirming that SOON scales linearly with H . Appendix C Theoretical Proofs C.1 Proof of Theorem 3.1 : Local Truncation Error of SOON Block Theorem 3.1 . Let \mathcal{L}_{\mathcal{Z}} and \mathcal{L}_{\mathcal{M}} denote the infinitesimal generators corresponding to the Zonal and Meridional operators, respectively. Let \mathcal{S}_{\mathrm{SOON}}(\tau) denote the evolution operator of one SOON block approximating the exact dynamics e^{\tau(\mathcal{L}_{\mathcal{Z}}+\mathcal{L}_{\mathcal{M}})} . By enforcing the symmetric composition \mathcal{Z}\circ\mathcal{M}\circ\mathcal{Z} (implemented as the Strang splitting e^{\frac{\tau}{2}\mathcal{L}_{\mathcal{Z}}}e^{\tau\mathcal{L}_{\mathcal{M}}}e^{\frac{\tau}{2}\mathcal{L}_{\mathcal{Z}}} ), the local truncation error is of third order, i.e., \|\mathcal{S}_{\mathrm{SOON}}(\tau)-e^{\tau(\mathcal{L}_{\mathcal{Z}}+\mathcal{L}_{\mathcal{M}})}\|=\mathcal{O}(\tau^{3}) , which is strictly superior to the \mathcal{O}(\tau^{2}) error of standard sequential stacking based on the Lie-Trotter splitting ( 39 ) . Proof. Let \mathcal{L}_{\text{total}}=\mathcal{L}_{\mathcal{Z}}+\mathcal{L}_{\mathcal{M}} . The exact evolution operator is \mathcal{E}(\tau)=\exp(\tau\mathcal{L}_{\text{total}}) . We analyze the local truncation error by comparing the logarithms of the numerical and exact operators via the Baker–Campbell–Hausdorff (BCH) formula. Sequential Stacking (Lie-Trotter Splitting). Standard architectures approximate the flow as \mathcal{S}_{\text{Seq}}(\tau)=e^{\tau\mathcal{L}_{\mathcal{M}}}e^{\tau\mathcal{L}_{\mathcal{Z}}} . Using the BCH expansion \ln(e^{X}e^{Y})=X+Y+\frac{1}{2}[X,Y]+\mathcal{O}(\|X,Y\|^{3}) with X=\tau\mathcal{L}_{\mathcal{M}} and Y=\tau\mathcal{L}_{\mathcal{Z}} : \displaystyle\ln(\mathcal{S}_{\text{Seq}}(\tau)) \displaystyle=\tau(\mathcal{L}_{\mathcal{M}}+\mathcal{L}_{\mathcal{Z}})+\frac{\tau^{2}}{2}[\mathcal{L}_{\mathcal{M}},\mathcal{L}_{\mathcal{Z}}]+\mathcal{O}(\tau^{3}) (11) \displaystyle=\ln(\mathcal{E}(\tau))+\frac{\tau^{2}}{2}[\mathcal{L}_{\mathcal{M}},\mathcal{L}_{\mathcal{Z}}]+\mathcal{O}(\tau^{3}). Since [\mathcal{L}_{\mathcal{M}},\mathcal{L}_{\mathcal{Z}}]\neq 0 in general, the dominant local error term is \mathcal{O}(\tau^{2}) . SOON Block (Symmetric Composition \mathcal{Z}\circ\mathcal{M}\circ\mathcal{Z} ). The SOON block implements the symmetric composition \mathcal{S}_{\text{SOON}}(\tau)=e^{\frac{\tau}{2}\mathcal{L}_{\mathcal{Z}}}e^{\tau\mathcal{L}_{\mathcal{M}}}e^{\frac{\tau}{2}\mathcal{L}_{\mathcal{Z}}} corresponding to \mathcal{Z}\circ\mathcal{M}\circ\mathcal{Z} . Applying the symmetric BCH formula ( 11 ) for \ln(e^{X/2}e^{Y}e^{X/2}) : \ln(e^{\frac{X}{2}}e^{Y}e^{\frac{X}{2}})=X+Y+\frac{1}{24}[Y,[Y,X]]-\frac{1}{12}[X,[X,Y]]+\mathcal{O}(\|X,Y\|^{4}). (12) Substituting X=\tau\mathcal{L}_{\mathcal{Z}} and Y=\tau\mathcal{L}_{\mathcal{M}} (both \mathcal{O}(\tau) ): \displaystyle\ln(\mathcal{S}_{\text{SOON}}(\tau)) \displaystyle=\tau(\mathcal{L}_{\mathcal{Z}}+\mathcal{L}_{\mathcal{M}})+\frac{\tau^{3}}{24}[\mathcal{L}_{\mathcal{M}},[\mathcal{L}_{\mathcal{M}},\mathcal{L}_{\mathcal{Z}}]]-\frac{\tau^{3}}{12}[\mathcal{L}_{\mathcal{Z}},[\mathcal{L}_{\mathcal{Z}},\mathcal{L}_{\mathcal{M}}]]+\mathcal{O}(\tau^{4}) (13) \displaystyle=\ln(\mathcal{E}(\tau))+\mathcal{O}(\tau^{3}). Crucially, the \mathcal{O}(\tau^{2}) commutator term [\mathcal{L}_{\mathcal{Z}},\mathcal{L}_{\mathcal{M}}] vanishes due to the symmetry of the composition \mathcal{Z}\circ\mathcal{M}\circ\mathcal{Z} . Error Bound. For sufficiently small \tau and under standard regularity assumptions on \mathcal{L}_{\mathcal{Z}},\mathcal{L}_{\mathcal{M}} (e.g., generating analytic semigroups on a Banach space), the exponential map satisfies \|e^{A}-e^{B}\|\leq C\|A-B\| for some constant C>0 independent of \tau ( 11 ) . Setting A=\ln(\mathcal{S}_{\text{SOON}}(\tau)) and B=\tau\mathcal{L}_{\text{total}} , we obtain: \|\mathcal{S}_{\text{SOON}}(\tau)-\mathcal{E}(\tau)\|\leq C\|\ln(\mathcal{S}_{\text{SOON}}(\tau))-\ln(\mathcal{E}(\tau))\|=\mathcal{O}(\tau^{3}). (14) Thus the SOON block achieves third-order local truncation error, strictly superior to the second-order local error of sequential stacking in prior global S2S forecasting works ( 19 ; 20 ; 21 ; 17 ) . ∎ Remark C.1 ( Significance in S2S Forecasting ) . The third-order local truncation error ( \mathcal{O}(\tau^{3}) ) of SOON yields second-order global accuracy ( \mathcal{O}(\tau^{2}) ), whereas Lie-Trotter splitting yields only first-order global accuracy ( \mathcal{O}(\tau) ). For Subseasonal-to-Seasonal forecasting over total time T with step size \tau (requiring N=T/\tau steps), the accumulated global error for Lie-Trotter splitting scales as N\cdot\mathcal{O}(\tau^{2})=\mathcal{O}(\tau) , while for SOON it scales as N\cdot\mathcal{O}(\tau^{3})=\mathcal{O}(\tau^{2}) . This quadratic reduction in global error significantly delays numerical divergence from the true atmospheric manifold, which is critical for capturing chaotic dynamics and preserving slow-varying climatic patterns (e.g., MJO) deep into the predictability desert. C.2 Proof of Proposition 3.2 : Spectral Distortion of LayerNorm Proposition 3.2 . Let \mathbf{z}\in\mathbb{R}^{C} be the latent representation of a latitudinal ring token in \bm{E}^{(0)} and let \hat{\mathbf{z}}=\mathcal{F}(\mathbf{z}) be its discrete Fourier spectrum. LayerNorm introduces spectral distortion by eliminating the zero-frequency component (background energy), i.e., |\mathcal{F}(\text{LayerNorm}(\mathbf{z}))_{0}|\equiv 0 , where \mathcal{F}(\cdot)_{0} represents the 0 -th frequency component of the DFT. Proof. Let \mathbf{z}=[z_{0},\dots,z_{C-1}]^{\top}\in\mathbb{R}^{C} . The Discrete Fourier Transform (DFT) is defined as \hat{z}_{k}=\sum_{n=0}^{C-1}z_{n}e^{-i2\pi kn/C} , where k=0 corresponds to the zero-frequency (DC) component. Layer Normalization centers the input via \mathbf{z}^{\prime}=\mathbf{z}-\mu\mathbf{1} , where \mu=\frac{1}{C}\sum_{n=0}^{C-1}z_{n} is the mean value. By the linearity of the DFT operator \mathcal{F} , the zero-frequency component of the centered signal is: \mathcal{F}(\mathbf{z}^{\prime})_{0}=\sum_{n=0}^{C-1}(z_{n}-\mu)=\sum_{n=0}^{C-1}z_{n}-C\cdot\left(\frac{1}{C}\sum_{n=0}^{C-1}z_{n}\right)=\hat{z}_{0}-\hat{z}_{0}=0. (15) The subsequent scaling by standard deviation \sigma in LayerNorm ( \frac{\mathbf{z}^{\prime}}{\sigma} ) involves multiplication by a scalar, which does not alter the zero value at k=0 . Thus, LayerNorm acts as a hard high-pass filter that strictly removes the DC component \hat{z}_{0} , which represents the background energy state of the atmospheric field. ∎ C.3 Proof of Proposition 3.3 : Spectral Fidelity of RMSNorm Proposition 3.3 . RMSNorm preserves the phase angles \arg(\mathcal{F}(\text{RMSN}(\mathbf{z}))_{k})=\arg(\hat{\mathbf{z}}_{k}) for all wave numbers k . Proof. RMSNorm applies a multiplicative scaling: \text{RMSN}(\mathbf{z})=\lambda\mathbf{z} , where \lambda=\big(\frac{1}{C}\sum_{n=0}^{C-1}z_{n}^{2}\big)^{-1/2} . Since the energy must be non-negative, \lambda is a strictly positive real scalar ( \lambda\in\mathbb{R}^{+} ). To analyze its spectral effect, we first invoke Parseval’s Theorem ( 29 ) to relate spatial-domain energy to spectral-domain energy. Derivation of Parseval’s Identity. Using the inverse DFT definition z_{n}=\frac{1}{C}\sum_{k=0}^{C-1}\hat{z}_{k}e^{i\frac{2\pi kn}{C}} : \displaystyle\sum_{n=0}^{C-1}z_{n}^{2} \displaystyle=\sum_{n=0}^{C-1}z_{n}z_{n}^{*}=\sum_{n=0}^{C-1}\left(\frac{1}{C}\sum_{k=0}^{C-1}\hat{z}_{k}e^{i\frac{2\pi kn}{C}}\right)\left(\frac{1}{C}\sum_{k^{\prime}=0}^{C-1}\hat{z}_{k^{\prime}}^{*}e^{-i\frac{2\pi k^{\prime}n}{C}}\right) (16) \displaystyle=\frac{1}{C^{2}}\sum_{k=0}^{C-1}\sum_{k^{\prime}=0}^{C-1}\hat{z}_{k}\hat{z}_{k^{\prime}}^{*}\underbrace{\sum_{n=0}^{C-1}e^{i\frac{2\pi(k-k^{\prime})n}{C}}}_{C\cdot\delta_{k,k^{\prime}}}=\frac{1}{C}\sum_{k=0}^{C-1}|\hat{z}_{k}|^{2}. This establishes that the total spatial energy is proportional to the total spectral energy. Substituting this into the RMSNorm scaling factor: \lambda=\left(\frac{1}{C}\sum_{n=0}^{C-1}z_{n}^{2}\right)^{-1/2}=\left(\frac{1}{C^{2}}\sum_{k=0}^{C-1}|\hat{z}_{k}|^{2}\right)^{-1/2}. (17) The spectral representation of the normalized signal is given by the linearity of the Fourier transform: \mathcal{F}(\text{RMSN}(\mathbf{z}))_{k}=\mathcal{F}(\lambda\mathbf{z})_{k}=\lambda\hat{z}_{k}. (18) Since \lambda is a positive real number, it affects only the magnitude of the complex spectral coefficients \hat{z}_{k} , but not their argument (phase angle). Therefore, for all wavenumbers k : \arg(\mathcal{F}(\text{RMSN}(\mathbf{z}))_{k})=\arg(\lambda\hat{z}_{k})=\arg(\hat{z}_{k}). (19) Furthermore, the relative spectral energy distribution is preserved: \frac{|\mathcal{F}(\text{RMSN}(\mathbf{z}))_{k}|^{2}}{|\mathcal{F}(\text{RMSN}(\mathbf{z}))_{k^{\prime}}|^{2}}=\frac{\lambda^{2}|\hat{z}_{k}|^{2}}{\lambda^{2}|\hat{z}_{k^{\prime}}|^{2}}=\frac{|\hat{z}_{k}|^{2}}{|\hat{z}_{k^{\prime}}|^{2}},\quad\forall k,k^{\prime}. (20) Critically, unlike LayerNorm, the zero-frequency component is preserved with non-zero magnitude ( |\mathcal{F}(\text{RMSN}(\mathbf{z}))_{0}|=\lambda|\hat{z}_{0}|>0 when \hat{z}_{0}\neq 0 ), maintaining the background energy state essential for geostrophic balance. ∎ Remark C.2 ( Physical Implication for S2S Forecasting ) . In atmospheric dynamics, the zero-frequency component ( \hat{z}_{0} ) corresponds to the planetary background flow energy (e.g., mean zonal kinetic energy), while higher wavenumbers represent eddy energy. LayerNorm’s elimination of the background energy disrupts the energy partitioning between mean flow and transient eddies, violating fundamental conservation principles. RMSNorm’s uniform spectral scaling preserves both the phase structure and relative energy distribution, including the critical background energy state, ensuring physically consistent energy cascades during long-horizon integration. This spectral fidelity prevents artificial energy leakage between scales, which is essential for maintaining correct amplitude evolution of planetary waves and mitigating forecast drift in the predictability desert. Appendix D Experimental Details D.1 Dataset We evaluate SOON on the ERA5 reanalysis dataset ( 13 ) , which provides a comprehensive record of global atmospheric conditions. To construct the training and evaluation data for our S2S forecasting task, we downsample the original 0.25^{\circ} ERA5 data to a coarser resolution of 1.5^{\circ}\times 1.5^{\circ} by sub-sampling grid points that align with the target grid. Specifically, we extract coordinates (\varphi,\lambda) on the latitude–longitude domain \Omega=[-90^{\circ},90^{\circ}]\times[-180^{\circ},180^{\circ}) , where \varphi\in\{-90^{\circ},-88.5^{\circ},\ldots,90^{\circ}\} and \lambda\in\{-180^{\circ},-178.5^{\circ},\ldots,178.5^{\circ}\} , resulting in a spatial resolution of 121\times 240 . This regridding strategy ensures that our input data structure is consistent with the outputs of baseline models such as PanguWeather ( 3 ) and FourCastNetV2 ( 30 ) when evaluated at this resolution. We select a set of 63 variables critical for capturing atmospheric dynamics, covering the time period from 1979 to 2018 with a temporal resolution of 1 day (daily averages). A detailed summary of the dataset configuration is provided in Table 3 . Table 3: Detailed configuration of the ERA5 dataset used in this work. Category Variable Name Symbol Levels Resolution Lat-Lon Range Time Pressure Level Geopotential z 10 1.5^{\circ} -90^{\circ}\text{S}\;180^{\circ}\text{W}\sim 90^{\circ}\text{N}\;180^{\circ}\text{E} 1979\sim 2018 Specific Humidity q 10 1.5^{\circ} -90^{\circ}\text{S}\;180^{\circ}\text{W}\sim 90^{\circ}\text{N}\;180^{\circ}\text{E} 1979\sim 2018 Temperature t 10 1.5^{\circ} -90^{\circ}\text{S}\;180^{\circ}\text{W}\sim 90^{\circ}\text{N}\;180^{\circ}\text{E} 1979\sim 2018 U Component of Wind u 10 1.5^{\circ} -90^{\circ}\text{S}\;180^{\circ}\text{W}\sim 90^{\circ}\text{N}\;180^{\circ}\text{E} 1979\sim 2018 V Component of Wind v 10 1.5^{\circ} -90^{\circ}\text{S}\;180^{\circ}\text{W}\sim 90^{\circ}\text{N}\;180^{\circ}\text{E} 1979\sim 2018 Vertical Velocity w 10 1.5^{\circ} -90^{\circ}\text{S}\;180^{\circ}\text{W}\sim 90^{\circ}\text{N}\;180^{\circ}\text{E} 1979\sim 2018 Single Level 2 Metre Temperature t2m 1 1.5^{\circ} -90^{\circ}\text{S}\;180^{\circ}\text{W}\sim 90^{\circ}\text{N}\;180^{\circ}\text{E} 1979\sim 2018 10 Metre U Wind u10 1 1.5^{\circ} -90^{\circ}\text{S}\;180^{\circ}\text{W}\sim 90^{\circ}\text{N}\;180^{\circ}\text{E} 1979\sim 2018 10 Metre V Wind v10 1 1.5^{\circ} -90^{\circ}\text{S}\;180^{\circ}\text{W}\sim 90^{\circ}\text{N}\;180^{\circ}\text{E} 1979\sim 2018 D.2 Baselines To rigorously evaluate the performance of SOON, we benchmark it against a comprehensive set of baseline models: Operational Forecasting Systems. These established numerical weather prediction (NWP) systems represent the current gold standard in operational meteorological forecasting. • CMA: The China Meteorological Administration implements the Beijing Climate Center’s fully-coupled BCC-CSM2-HR model ( 46 ) , providing operational forecasts with a 60-day lead time. • NCEP: The National Centers for Environmental Prediction deploys the Climate Forecast System version 2 (CFSv2) ( 35 ) to generate daily control forecasts extending to a 45-day prediction window. • ECMWF: The European Centre for Medium-Range Weather Forecasts employs the Integrated Forecasting System (IFS) ( 24 ) . We utilize the control forecasts from cycle 47r3, which are widely regarded as the most skillful operational baseline for S2S prediction. Data-Driven Models. We compare against both pre-trained foundation models and models trained from scratch. • PanguWeather ( 3 ) : A 3D Earth-specific Transformer that tokenizes pressure-level and surface data separately, trained using a hierarchical temporal aggregation strategy. We use the API https://github.com/ecmwf-lab/ai-models to perform inference. • FourCastNetV2 (FCN2) ( 30 ) : An iterative model based on Vision Transformers that incorporates Adaptive Fourier Neural Operators (AFNO) for efficient spatial token mixing. We use the API https://github.com/ecmwf-lab/ai-models to perform inference. • Transformer ( 41 ) : The standard self-attention architecture adapted for global grid inputs, serving as a fundamental baseline for sequence modeling capability. • FNO ( 18 ) : The Fourier Neural Operator, which learns resolution-invariant operators by parameterizing the integral kernel in the frequency domain, adapted here for global forecasting. • ViT ( 9 ) : The standard Vision Transformer employing a naive grid patching strategy, treating global weather fields as flat images without specific geometric adaptations. • ClimaX ( 27 ) : A foundation model for weather and climate using a Vision Transformer backbone with variable-specific tokenization and aggregation, retrained on our dataset. • ClimODE ( 42 ) : A physics-informed approach that models atmospheric dynamics using Neural Ordinary Differential Equations (ODEs) to capture continuous-time evolution. • CirT ( 19 ) : A Transformer-based model designed for S2S forecasting that introduces circular patching and frequency-domain mixing to address spherical boundary conditions. • EIMP ( 20 ) : A graph-based model utilizing equivariant and invariant message passing to strictly enforce rotational symmetries on spherical grids. • TelePiT ( 21 ) : A physics-informed Transformer that integrates spherical harmonic embeddings with neural ODEs to explicitly capture multi-scale teleconnections. D.3 Implementation Details The SOON model is implemented using PyTorch 2.3.1 and trained on 8 NVIDIA GPUs via Distributed Data Parallel (DDP) with mixed precision (FP16), while FFT operations within the Zonal Operator are forced to FP32 to ensure numerical stability. We utilize an anisotropic embedding kernel size of (1,240) and a meridional spatial kernel size of k=7 . The model is trained for 20 epochs with a global batch size of 256 using the AdamW optimizer (weight decay 10^{-5} , \epsilon=10^{-8} ) and gradient clipping at 1.0. The learning rate follows a cosine annealing schedule, decaying from a peak of 10^{-3} to 10^{-4} . To determine the optimal architecture, we conducted a grid search on the validation set over the following hyperparameter spaces: embedding dimension C\in\{128,192,256,384,512\} , network depth L\in\{4,5,6,7,8,9\} , decoder depth \in\{1,2,3\} , learning rate \in\{5\times 10^{-4},1\times 10^{-3},5\times 10^{-3}\} , MLP ratio \in\{2.0,4.0,6.0\} , and dropout rates \in\{0.0,0.1,0.2\} . The final selected configuration is C=256 , L=7 , decoder depth 1, learning rate 10^{-3} , MLP ratio 4.0, and dropout 0.1, resulting in a parameter count of approximately 16.58M. We have also conducted comprehensive ablation studies as aforementioned in Section 4.3 . The variants are implemented as follows: 1. Variant without Anisotropic Embedding (w/o AE): We replace the proposed (1,W) anisotropic convolution with a standard isotropic patch embedding of size 4\times 4 , which flattens the input into a generic token sequence devoid of latitudinal ring topology. 2. Variant without Zonal Operator (w/o ZO): We remove the spectral mechanism and replace the \mathcal{Z}\circ\mathcal{M}\circ\mathcal{Z} symmetric structure with stacked Meridional operators ( \mathcal{M}\circ\mathcal{M} ) to maintain model capacity. 3. Variant without Meridional Operator (w/o MO): We replace the spatial operator with a double Zonal structure ( \mathcal{Z}\circ\mathcal{Z} ). 4. Variant with unshared-weight Zonal Operator (w/ unshared ZO): We define two independent Zonal Operators, \mathcal{Z}_{1} and \mathcal{Z}_{2} , resulting in an asymmetric structure \mathcal{Z}_{1}\circ\mathcal{M}\circ\mathcal{Z}_{2} . 5. Variant with LayerNorm (w/ LN): We replace all RMSNorm layers with standard Layer Normalization to evaluate the impact of mean-centering on long-term stability. Appendix E Additional Results E.1 Performance w.r.t. the number of SOON blocks We evaluate the impact of network depth by varying the number of SOON blocks L\in\{4,5,6,7,8,9\} . As can be seen from Figure 7 , forecasting performance consistently improves as depth increases, peaking at L=7 where RMSE reaches a minimum for key variables. Increasing the depth further yields diminishing returns or slight degradation. Thus, we select L=7 to balance accuracy and computational efficiency. Figure 7: RMSE and ACC variations across different numbers of SOON blocks ( L ) indicate that L=7 achieves the best trade-off between performance and model complexity. E.2 Spectral Energy Distribution Analysis Figure 8: SpecRes comparison across pressure levels. SpecRes measures spectral energy distribution accuracy (lower is better; lighter colors indicate smaller values). We further assess spectral energy distribution accuracy using SpecRes ( 21 ) across all pressure levels for the prognostic variables z , t , u , and v under weeks 3–4 and weeks 5–6 forecasts. SpecRes quantifies the root of the expected squared residual in the spectral domain, i.e., \text{SpecRes}=\sqrt{\mathbb{E}_{k}\!\left[(\hat{S}^{\prime}(k)-S^{\prime}(k))^{2}\right]} , where \hat{S}^{\prime}(k) and S^{\prime}(k) denote the normalized power spectra of prediction and target, respectively. Therefore, lower SpecRes values indicate a more accurate preservation of spectral energy distribution, especially for high-frequency wavenumbers. As shown in Figure 8 , SOON consistently exhibits substantially lower SpecRes values than all competing models over nearly all vertical levels, indicating a more faithful preservation of spectral energy distribution during long-horizon integration. This empirical observation aligns with our theoretical analysis: compared with mean-centering normalization, RMSNorm better preserves phase stability and background energy, thereby mitigating long-term spectral drift and improving spectral fidelity. E.3 Additional Global Visualizations We provide expanded global error maps to visually assess spatial performance distributions across different variables and lead times. Figure 9 through Figure 15 display the RMSE distributions for t850,t500,z500, and z850 during weeks 3–4 and weeks 5–6. Consistent with the findings in the main text, SOON exhibits deep blue patterns globally, indicating low error. A recurrent weakness in baseline models is the high error accumulation in polar regions (red/orange zones), caused by the projection distortion of isotropic kernels on the sphere. SOON effectively eliminates these artifacts, maintaining high accuracy in both the tropics and high latitudes. Furthermore, even as the forecast horizon extends to weeks 5–6, SOON preserves structural details better than competitors, which tend to produce overly smooth or erroneous fields. Figure 9: Global RMSE distribution for t850 at weeks 5–6 lead time. Figure 10: Global RMSE distribution for t500 at weeks 3–4 lead time. Figure 11: Global RMSE distribution for t500 at weeks 5–6 lead time. Figure 12: Global RMSE distribution for z500 at weeks 3–4 lead time. Figure 13: Global RMSE distribution for z500 at weeks 5–6 lead time. Figure 14: Global RMSE distribution for z850 at weeks 3–4 lead time. Figure 15: Global RMSE distribution for z850 at weeks 5–6 lead time. E.4 Vertical Profile Analysis We further analyze the model’s performance across the vertical atmospheric column to ensure physical consistency at different altitudes. Figure 16 compares the RMSE of wind components ( u,v ) against data-driven baselines, while Figure 17 benchmarks geopotential ( z ) and temperature ( t ) against operational NWP systems. As shown in the heatmaps, SOON achieves the lowest normalized RMSE (indicated by the lightest colors) across almost all 13 pressure levels, from the surface (1000 hPa) up to the stratosphere (50 hPa). This vertical consistency is particularly notable in comparison to models like ClimaX and EIMP, which exhibit degradation in the upper atmosphere. The results demonstrate that SOON’s anisotropic embedding and operator design effectively capture the distinct dynamics of different atmospheric layers without requiring level-specific fine-tuning. Figure 16: RMSE comparison between SOON and data-driven models on wind variables u and v across different pressure levels. Figure 17: RMSE comparison between SOON and operational NWP systems on variables z and t across different pressure levels. E.5 Daily Visualization and Temporal Stability To probe fine-grained temporal stability, we trained SOON and other baselines to output the full daily trajectory. Figure 18 and Figure 19 visualize the day-by-day evolution of RMSE and ACC for weeks 3–4 (days 15–28) and weeks 5–6 (days 29–42), respectively. In the weeks 3–4 window (Figure 18 ), SOON consistently maintains superior performance across all variables. While baselines suffer from rapid degradation over time (manifested as steep error slopes), SOON exhibits a remarkably flat error growth rate. This advantage becomes even more pronounced in the extended weeks 5–6 window (Figure 19 ). For highly chaotic surface variables like v10 , the performance gap between SOON and the nearest competitor widens significantly as the lead time increases. This empirical evidence confirms that our symmetric operator splitting scheme effectively mitigates the error accumulation inherent in long-horizon forecasting, maintaining high correlation (ACC) even deep into the subseasonal range. Figure 18: Daily RMSE and ACC variations across critical pressure level and single level variables during weeks 3–4 (Days 15–28). Figure 19: Daily RMSE and ACC variations across critical pressure level and single level variables during weeks 5–6 (Days 29–42). E.6 Seasonal Robustness Analysis S2S forecasting skill often varies significantly by season due to changing atmospheric dynamics. To evaluate robustness, we break down the testing set performance by month. Figure 20 to Figure 23 illustrate the monthly RMSE for key variables ( z500,z850,t500,t850 ). SOON demonstrates consistent superiority across all months of the year, outperforming both data-driven and numerical baselines. Whether in Northern Hemisphere winter (where baroclinic instability is high) or summer, SOON maintains the lowest error margins. This suggests that the model has successfully learned generalized representations of seasonal cycle dynamics rather than overfitting to specific regimes. Figure 20: Monthly RMSE breakdown for z500 : SOON consistently outperforms other models across all seasons. Figure 21: Monthly RMSE breakdown for z850 : SOON consistently outperforms other models across all seasons. Figure 22: Monthly RMSE breakdown for t500 : SOON consistently outperforms other models across all seasons. Figure 23: Monthly RMSE breakdown for t850 : SOON consistently outperforms other models across all seasons. Appendix F Limitations While SOON achieves state-of-the-art performance, it currently relies on a deterministic regression objective, which does not explicitly quantify the forecast uncertainty inherent in chaotic S2S timescales. Additionally, due to computational constraints, our experiments are conducted at a 1.5^{\circ} resolution, potentially under-resolving fine-grained sub-grid processes. Future extensions could incorporate probabilistic frameworks and higher-resolution training to further enhance utility for operational decision-making.