🌍 Earth Science

🔬 ICLR2026 · 7 paper notes

📌 Same area in other venues: 📷 CVPR2026 (1) · 🧪 ICML2026 (2) · 🤖 AAAI2026 (2) · 🧠 NeurIPS2025 (6)

GeoFAR: Geography-Informed Frequency-Aware Super-Resolution for Climate Data

GeoFAR decomposes the low-frequency bias in climate super-resolution into two problems: "under-represented frequency components" and "missing geographical conditions." It utilizes DCT frequency convolutional kernels to extract fine-grained frequency band representations and modulates these representations pixel-wise using a geographic implicit representation (Geo-INR) composed of longitude, latitude, and elevation. This approach significantly reduces high-frequency errors and prediction biases in complex terrain across multi-scale climate downscaling tasks such as ERA5, PRISM, and CERRA.

OmniField: Conditioned Neural Fields for Robust Multimodal Spatiotemporal Learning

OmniField models "scientific observation data" (climate, air pollution) as a continuous neural field conditioned on available modalities. Utilizing Multimodal Crosstalk (MCT) blocks and Iterative Cross-Modal Refinement (ICMR), it aligns heterogeneous signals before decoding. This unified framework supports reconstruction, interpolation, and prediction without gridding or interpolation preprocessing, reducing average error by 22.4% relative to 8 strong baselines while maintaining performance under severe sensor noise.

RainPro-8: An Efficient Deep Learning Model to Estimate Rainfall Probabilities Over 8 Hours

RainPro-8 utilizes a MaxViT-U-Net with only 36.7M parameters to fuse multi-source data from radar, satellite, and Numerical Weather Prediction (NWP). Through "ordered consistent loss + single-forward prediction," it outputs high-resolution probabilistic precipitation forecasts for Europe over 8 hours in one go. It outperforms existing NWP, extrapolation, and deep learning nowcasting methods while being 48x faster in inference than MetNet-like models.

Task-Adaptive Parameter-Efficient Fine-Tuning for Weather Foundation Models

For Weather Foundation Model (WFM) fine-tuning, this paper proposes WeatherPEFT: the forward pass uses Task-Adaptive Dynamic Prompting (TADP) to extract "variable × resolution × spatiotemporal" task features from encoder embedding weights to generate soft prompts, while the backward pass employs Stochastic Fisher-Guided Adaptive Selection (SFAS) to update only a small subset of parameters with the highest Fisher information. It matches or exceeds Full-Tuning on three downstream tasks using only ~0.3%–4% trainable parameters.

The Seismic Wavefield Common Task Framework

This paper adapts the "Common Task Framework (CTF)" approach—which catalyzed benchmarks like ImageNet and AlphaZero in NLP/CV—to seismology. It provides three multi-scale seismic wavefield datasets alongside a 12-point scoring protocol using hidden test sets. Evaluating 18 mainstream scientific machine learning models reveals that most complex architectures fail to outperform a naive "all-zero" baseline.

TianQuan-S2S: Constructing Subseasonal-to-Seasonal Global Weather Forecasting Models by Incorporating Climatology

TianQuan-S2S integrates "long-term climatological means" into patch embeddings via attention fusion and injects learnable Gaussian noise into each layer of a ViT. This specifically addresses the "model collapse" (increasingly blurry predictions) of data-driven models in 15–45 day subseasonal forecasting, outperforming both the numerical model ECMWF-S2S and the data-driven FuXi-S2S on the ERA5 dataset.

Uncovering the Mechanism of Continuous Representation Full Waveform Inversion: A Wave-based Neural Tangent Kernel Framework

This paper extends Neural Tangent Kernel (NTK) theory to Full Waveform Inversion (FWI), proposing a "Wave-based NTK" to unify the characterization of traditional FWI and Continuous Representation FWI (CR-FWI). It explains the phenomenon "why INR representations are more robust but converge slowly at high frequencies" through eigenvalue decay rates. Based on this, it designs IG-FWI, a hybrid of INR and multi-resolution grids, achieving a superior trade-off between robustness and convergence speed.