🌍 Earth Science

🧠 NeurIPS2025 · 5 paper notes

A Probabilistic U-Net Approach to Downscaling Climate Simulations

This work presents the first application of a probabilistic U-Net to statistical climate downscaling (16× super-resolution). By sampling from a variational latent space, the model generates ensemble forecasts for uncertainty quantification. The paper systematically compares four training objectives — WMSE, MS-SSIM, WMSE-MS-SSIM, and afCRPS — revealing complementary trade-offs between extreme event capture and fine-scale spatial variability preservation.

Adaptive Online Emulation for Accelerating Complex Physical Simulations

This paper proposes Adaptive Online Emulation (AOE), a framework that dynamically trains an ELM-based neural network surrogate model during physical simulation execution to replace expensive computational components, requiring no offline pretraining. On an exoplanetary atmospheric simulation, AOE achieves an 11.1× speedup (91% time savings) with only ~0.01% accuracy loss.

ControlFusion: A Controllable Image Fusion Framework with Language-Vision Degradation Prompts

This paper proposes ControlFusion, a controllable infrared-visible image fusion framework based on language-vision degradation prompts. It employs a physics-driven degradation imaging model to simulate compound degradations, and uses a prompt-modulated network to perform dynamic restoration and fusion, achieving comprehensive state-of-the-art performance under both real-world and compound degradation scenarios.

Predicting Public Health Impacts of Electricity Usage

This paper proposes HealthPredictor, an AI pipeline that maps electricity consumption end-to-end to public health damages (measured in $/MWh), comprising three modules: fuel mix prediction, air quality conversion, and health impact assessment. Health-driven optimization significantly reduces health impact prediction error compared to fuel-mix-driven baselines, and achieves a 24–42% reduction in health damages in an EV charging scheduling case study.

Reasoning With a Star: A Heliophysics Dataset and Benchmark for Agentic Scientific Reasoning

This paper introduces Reasoning With a Star (RWS), a 158-question scientific reasoning benchmark derived from NASA heliophysics summer school problem sets, spanning three answer types (numeric/symbolic/textual). Paired with a unit-aware grader, it evaluates four multi-agent coordination paradigms (HMAW/PACE/PHASE/SCHEMA) and finds that no single paradigm dominates across all tasks — the systems-engineering-inspired SCHEMA achieves the strongest performance on tasks requiring rigorous constraint validation.