KunPeng: A Global Ocean Environmental Model

TL;DR

KunPeng tackles global ocean environmental prediction by transferring meteorological large-model techniques to the ocean domain, addressing land–sea discontinuities, multi-scale spatial dynamics, and slow-varying temporal evolution. The model combines a terrain-adaptive masked loss (MLL1), a longitude-cyclic deformable convolution (LC-DCN), and a deformable convolution-enhanced multi-step predictor (DC-MTP) to form a multi-scale, temporally coherent predictor. Results show an average 15-day ACC of $0.80$ at $0.25^{\circ}$ resolution with notable reductions in $MSE$ and $MAE$ relative to baselines, and strong performance in SST, deep-sea, and current velocity forecasts. The work also discusses cross-domain applicability, land–sea boundary dynamics, and future directions, including meteorological coupling data, long-period feature capture, and ensemble approaches, highlighting potential for high-resolution, coupled ocean–atmosphere forecasting.

Abstract

Inspired by the similarity of the atmosphere-ocean physical coupling mechanism, this study innovatively migrates meteorological large-model techniques to the ocean domain, constructing the KunPeng global ocean environmental prediction model. Aimed at the discontinuous characteristics of marine space, we propose a terrain-adaptive mask constraint mechanism to mitigate effectively training divergence caused by abrupt gradients at land-sea boundaries. To fully integrate far-, medium-, and close-range marine features, a longitude-cyclic deformable convolution network (LC-DCN) is employed to enhance the dynamic receptive field, achieving refined modeling of multi-scale oceanic characteristics. A Deformable Convolution-enhanced Multi-Step Prediction module (DC-MTP) is employed to strengthen temporal dependency feature extraction capabilities. Experimental results demonstrate that this model achieves an average ACC of 0.80 in 15-day global predictions at 0.25$^\circ$ resolution, outperforming comparative models by 0.01-0.08. The average mean squared error (MSE) is 0.41 (representing a 5%-31% reduction) and the average mean absolute error (MAE) is 0.44 (0.6%-21% reduction) compared to other models. Significant improvements are particularly observed in sea surface parameter prediction, deep-sea region characterization, and current velocity field forecasting. Through a horizontal comparison of the applicability of operators at different scales in the marine domain, this study reveals that local operators significantly outperform global operators under slow-varying oceanic processes, demonstrating the effectiveness of dynamic feature pyramid representations in predicting marine physical parameters.

Figures (25)

• Figure 1: Architecture of the KunPeng Ocean Large Model. a) is basic model structure. The input consists of three-dimensional ocean environmental parameters, two-dimensional ocean environmental parameters, meteorological environmental parameters at time T, and the static field, which is stacked along the coastal layer channels to form a ($C$,$H$,$W$)-dimensional tensor. After passing through the Embedding layer, a ($C_1$,$H/4$,$W/4$)-dimensional hidden-layer feature is obtained. The feature mixing part is composed of an Encoder and a Decoder, each of which contains two operator blocks, and each block consists of four layers with the same structure. In the Encoder, the two blocks are connected by a down-sampling layer. Through a convolution with a stride of 2, the hidden layer is sampled into a ($2C_1$,$H/8$,$W/8$)-dimensional tensor. The up-sampling layer of the Decoder uses a transposed convolution with the same stride to perform the opposite operation. Finally, the Recovery layer is used to restore the block and reconstruct the ($C_2$,$H$,$W$)-dimensional atmospheric/oceanic physical field at time T+1. Generally, $C_2 \leq C_{1}$ to remove the static fields that do not need to be predicted, such as the ocean layer depth. To ensure numerical stability, there is a residual connection between the Embedding and Recovery layers. b) is Single-layer Structure in the Operator Block. Here, Norm represents layer normalization, FFN is the feed-forward neural network, which is mainly used for feature channel mixing, and LC-DCN is the longitude-cyclic deformable convolution, which is mainly used for feature space mixing. Each mixer has a residual link. c) is Schematic Diagram of the Longitude-cyclic Deformable Convolution. The double-circled points represent the pixel positions for the convolution calculation, and the squares represent the positions of the convolution kernel units after being adjusted by the deformation offset. Some units of the same convolution kernel can cross the left-right boundaries to perform the longitude-cyclic convolution operation. d) is Schematic Diagram of the DC-MTP(Deformable Convolution-enhanced Multi-step Prediction). The primary model's loss function $L_{main}$ and the auxiliary prediction model's loss function $L^{i}_{MTP}$ both employ Masked Dimension-Weighted L1 Loss (MLL1). The auxiliary model is only used during training to improve the main model's ability to capture temporal features and will be removed during inference.
• Figure 2: Comparison of selected model metrics. Here, $T$ denotes temperature, $S$ salinity, $U$ meridional current velocity, and $V$ zonal current velocity. Numerical labels indicate sea layer depth. Metrics were derived from daily 15-day forecast results spanning 2021-01-01T00:00:00Z UTC to 2021-12-16T00:00:00Z UTC. Lower values of MAE and MSE indicate better performance, with a theoretical minimum of 0, while higher ACC values signify improved accuracy, approaching a theoretical optimum of 1.
• Figure 3: This is a prediction map of the global ocean area at 2021-07-16T00:00:00Z UTC. This prediction was obtained through 15 cycles of data using the data at 2021-07-01T00:00:00Z UTC as the initial value. Among them, $T_1$, $S_1$, $U_1$, and $V_1$ represent the sea temperature at a depth of 1 meter, salinity, meridional current velocity, and zonal current velocity, respectively.
• Figure 4: This is a prediction map of the region between $10^\circ$S and $25^\circ$N in latitude between $100^\circ$W and $40^\circ$W in longitude at 2021-07-11T00:00:00Z UTC. This prediction was derived from 10 cycles of data based on the data at 2021-07-01T00:00:00Z UTC. Among them, $T1$, $S1$, $U1$, and $V1$ represent the sea temperature at a depth of 1 m, salinity, meridional current velocity, and zonal current velocity, respectively.
• Figure 5: Comparison of 15 - day Temperature Predictions