Building Machine Learning Limited Area Models: Kilometer-Scale Weather Forecasting in Realistic Settings

TL;DR

A framework for building kilometer-scale machine learning limited area models with boundary conditions imposed through a flexible boundary forcing method is presented, which achieves great potential for machine learning limited area models in the future of regional weather forecasting.

Abstract

Machine learning is revolutionizing global weather forecasting, with models that efficiently produce highly accurate forecasts. Apart from global forecasting there is also a large value in high-resolution regional weather forecasts, focusing on accurate simulations of the atmosphere for a limited area. Initial attempts have been made to use machine learning for such limited area scenarios, but these experiments do not consider realistic forecasting settings and do not investigate the many design choices involved. We present a framework for building kilometer-scale machine learning limited area models with boundary conditions imposed through a flexible boundary forcing method. This enables boundary conditions defined either from reanalysis or operational forecast data. Our approach employs specialized graph constructions with rectangular and triangular meshes, along with multi-step rollout training strategies to improve temporal consistency. We perform systematic evaluation of different design choices, including the boundary width, graph construction and boundary forcing integration. Models are evaluated across both a Danish and a Swiss domain, two regions that exhibit different orographical characteristics. Verification is performed against both gridded analysis data and in-situ observations, including a case study for the storm Ciara in February 2020. Both models achieve skillful predictions across a wide range of variables, with our Swiss model outperforming the numerical weather prediction baseline for key surface variables. With their substantially lower computational cost, our findings demonstrate great potential for machine learning limited area models in the future of regional weather forecasting.

Figures (49)

• Figure 1: Overview of the graph-based approach applied in our setting. The dual MLP encoders for interior and boundary inputs are introduced in this work.
• Figure 2: Desiderata for the boundary forcing framework.
• Figure 3: Illustration of different steps in the rectangular graph creation process. Nodes are laid out in a regular grid and connected to neighbors horizontally, vertically and diagonally. The set of mesh levels, with different number of nodes. Note that the vertical position of mesh levels is purely for visualization purposes and does not relate to vertical atmospheric levels. All mesh levels should be conceptually viewed as laying flat on the surface of the earth. 3-level hierarchical mesh graph, including edges connecting the different levels. Each node is connected to a set of $3 \times 3$ nodes at the level below, if present.
• Figure 4: Illustration of different steps in the triangular graph creation process. The global icosahedron that the subdividing process starts from. Triangle splitting method. Each triangle of 3 nodes and 3 edges is split into 6 nodes and 9 edges. Nodes and edges in triangular graph, subset to the convex hull of all grid points. 3-level hierarchical mesh graph, including edges connecting the different levels. Each node is connected to 7 nodes at the level below, if present.
• Figure 5: The regions being modeled in our two settings. The boundary region shown here has width 800km. Note that while the boundary area can be substantially larger than the interior, the spatial resolution of the boundary is much lower, so it consists of fewer grid points.