4DVarGen is proposed, a 4DVar-inspired generative framework for reconstructing sea surface variable fields at eddy-resolving scales from sparse remote-sensing observations that establishes a mathematical equivalence between 4DVar and an observation-guided denoising process.
This work offers a scalable pathway for next-generation Earth system models to learn directly from sparse, incomplete real-world observations and derives an optimization framework based on the expectation-maximization (EM) algorithm that enable learning directly from sparse and noisy observations.
Yangyang Kong, Yutong Jiang, Yanhai Gan et al.· 0 citations
A core inverse problem in the experimental sciences is the inference of a hidden dynamical state from sparse or indirect measurements. There is a natural opportunity for deep learning methods here, but machine-learnt reconstruction methods typically require full state data for training. We present Trajectory-Consistent Network Training (TraCTra), a label-free framework that trains reconstruction networks using only partial observation sequences and a differentiable forward model. TraCTra requires the network reconstruction and dynamical evolution to be mutually consistent: network-predicted states are marched forward in time to match subsequent observations and to agree in the full state space with independent reconstructions at later times. Across four fluid systems, the same objective reconstructs three-dimensional turbulence from coarse-grained fields, velocity from observations of density fluctuations, and three-dimensional density and velocity from sequences of projected two-dimensional shadowgraphs, while also recovering global vorticity from observations confined to a small spatial window. TraCTra outperforms assimilation-only and physics-informed neural approaches, preserves dynamically important multiscale structure, and remains accurate beyond the optimisation window. It transfers to held-out times in the three-dimensional shadowgraph problem and, when trained across trajectories, generalises to unseen flows in the two-dimensional problem. The results establish trajectory consistency as a general supervision principle for reconstructing hidden dynamical states without full state training targets.
The first oceanic 4D sparse observation reconstruction dataset, named OceanVerse, is presented, providing a novel large-scale dataset that meets the MNAR (Missing Not at Random) condition, supporting more effective model comparison, generalization evaluation and potential advancement of scientific reconstruction architectures.
Bin Lu, Jingjing Shen, Ze Zhao et al.· Proceedings of the 32nd ACM...· 0 citations
Accurate representation of atmospheric moisture is essential for reliable weather forecasting, particularly for small-scale convective systems and extreme events. However, determining high-resolution water vapor (WV) fields remains challenging. Conventional global navigation satellite system (GNSS) troposphere tomography reconstructs 4-D atmospheric wet refractivity fields but is limited by sparse and uneven ray paths, an ill-conditioned coefficient matrix, and an ill-posed inverse problem. Stabilization through constraints and regularization may introduce biases, while the low probability of ray–ray intersections in the lowest tropospheric layers reduces observational influence, causing some regions to depend more on background models than observations. To address these limitations, DeepTomo, to the best of the authors’ knowledge, the first artificial intelligence (AI)-based 4-D GNSS troposphere tomography is introduced as an explainable physics-informed deep learning approach that combines hybrid observational constraints with spatiotemporal learning. Beyond tomographic reconstruction, DeepTomo is conceived as an AI-based assimilation of GNSS observations into ERA5 fields; it integrates a 3-D convolutional neural network (CNN) with residual learning and attention mechanisms and employs a hybrid physics-informed loss function that combines GNSS-derived zenith wet delay (ZWD) with radio occultation (RO) and radiosonde refractivity profiles to correct the ERA5 background toward observational constraints. By learning spatiotemporal relationships between observations and background fields, DeepTomo refines wet refractivity estimates and enables physically consistent reconstruction even in voxels with limited observations. Trained and validated over a dense GNSS network in coastal California using a six-month dataset and evaluated against radiosonde and GNSS-derived ZWD data, DeepTomo performs strongly during the extreme weather event of Hurricane Hilary, a tropical cyclone (TC), in August 2023. Compared with conventional voxel-based tomography, it reduces the root mean square error (RMSE) by up to 64.85% during the TC and 41.62% overall, capturing large moisture variability. Explainable AI (XAI) analysis reveals dynamic spatial attention to regions of enhanced variability. A preliminary sensitivity analysis using GraphCast forecasts shows that the moisture corrections introduced by DeepTomo correspond to short-range forecast errors. This analysis provides an initial indication that DeepTomo, by producing physically consistent GNSS-constrained moisture analyses from ERA5 background fields, has the potential to improve initial conditions and forecast performance in next-generation AI weather forecasting systems, such as GraphCast, bridging GNSS observations with AI-based forecast initialization.
Saeid Haji-Aghajany, Benedikt Soja, Kefei Zhang et al.· IEEE Transactions on Geoscie...· 0 citations
Regional high-resolution ocean environmental forecasting combines spatial numerical modeling with temporal prediction, and is essential for monitoring the ecological security of specific ocean regions. In recent years, deep learning methods are generally more computationally efficient than traditional numerical models and enable fast, accurate forecasting. However, as data resolution increases, the training and computational costs of existing approaches increase substantially. To address this issue, we introduce Slow-OCast, a transfer-learning based model designed for high-resolution ocean environmental forecasting. Specifically, Slow-OCast incorporates the slow-varying motion characteristics of the ocean and comprises two insightful modules. The Fluid Motion Separator that injects low-frequency background dynamics into the fine-tuning process of a foundation model, functioning as a "magnifier" to encode physical priors of ocean dynamics. The Hydrokinetic Energy Path Integrator that provides an implicit representation of flow-field evolution, serving as a "compass" to guide accurate change prediction. We evaluate Slow-OCast on two high-resolution Mediterranean datasets, and results demonstrate Slow-OCast consistently outperforms all baseline methods across forecasting tasks with different lead times.
Qixiu Li, Xiang Zhu, Xiaoyong Li et al.· Proceedings of the 32nd ACM...· 0 citations
Deep learning models for scientific spatio-temporal downscaling often minimize reconstruction error while failing to preserve physically meaningful multi-scale structure. For sea surface temperature prediction, this can yield outputs that are numerically plausible yet overly smooth, missing mesoscale variability critical to regional ocean dynamics. Existing methods often focus on pixel-wise objectives or single-context conditioning, which limits their ability to preserve spectral fidelity and generalize across regions. To address this, we propose EddyFlow, a representation learning framework for kilometer-scale sea surface temperature downscaling that balances predictive accuracy, scale-dependent structure, and regional generalization. EddyFlow is trained on the Gulf of St.~Lawrence and evaluated in zero-shot and few-shot settings on the Bay of Fundy and the Gulf of Mexico. EddyFlow demonstrates that physics-informed representation learning reduces zero-shot RMSE by 21%, achieves up to 85.6% skill relative to persistence on unseen domains, and maintains near-ideal spectral fidelity with a PSD ratio of $\approx 1.00$.
Parth Doshi, Priyanka Aravindan, Vaishnav Vaidheeswaran et al.· 0 citations