To better understand hydrological processes and improve flood simulation, combining artificial intelligence (AI) with process-based models is a promising direction. We introduce a hybrid physics–AI approach that seamlessly integrates neural networks into a distributed hydrological model to refine water flow dynamics within an implicit numerical scheme. The hybrid models demonstrate strong performance and interpretable results, leading to reliable streamflow simulations for flood modeling.

Due to the limited spatial and temporal coverage of observations, global models are essential tools to study climate. However, long-term climate data at fine spatial scale are difficult to obtain because of elevated computational costs such algorithms involve. This paper presents a simple model based on the description of climate/topography interactions to generate local precipitation fields at low cost. The objective is to provide a flexible and easy to use method for paleoclimate studies.

By using the newly developed Python library JAX, we developed a fast and differentiable large-eddy simulation model, named LEX. Evaluated with a warm bubble case, LEX maintains high accuracy as the Cloud Model 1. With the hardware acceleration and better numerical stability, LEX can be quite faster. To report its differentiability, we further trained a deep learning-based parameterization scheme. The newly trained model can surpass the conventional scheme and get proper forecast results.

Drought is reshaping Europe's forests. Using an uncalibrated process-based model across 149 German sites, we identified key drivers of tree mortality in European beech and Norway spruce forests. Our model captured both the timing and extent of mortality. A new bark beetle module improved predictions for spruce. High soil water capacity and heterogeneous soils reduced drought impacts. These findings offer new insights to anticipate forest responses in a warming, drying climate.

Earth System Models are an essential tool for climate studies, yet temperature-sensitive parameters are often absent, resulting in a gap in model predictive capabilities. Organic carbon breakdown, also known as remineralization, is one such process. Here, we add this parameter to the Community Earth System Model and find improved regional patterns of carbon export. The new code will serve as a useful tool to improve the examination of marine carbon cycle feedbacks to changing climate conditions.

The new Western United States Large Forest Fire Stochastic Simulator (WULFFSS) is a monthly gridded model to simulate forest fires across the western United States in response to vegetation, topographic, anthropogenic, and climate factors. The model is highly skillful, accounting for over 80 % of the observed variability in annual forest-fire area and capturing observed spatial, intra-annual variations, and trends.