Ambient Solar Wind Speed Forecast with Physics-Informed Machine Learning

We present a novel physics-informed machine learning (ML) model designed to forecast the background (ambient) solar wind up to five days in advance. Solar wind speed is a critical driver of geomagnetic activity, and inaccuracies in its prediction significantly contribute to large errors in forecasting the arrival times of coronal mass ejections (CMEs), which are typically off by at least 10 hours.

Predicting solar wind speed has historically been a challenging task, with even state-of-the-art models often failing to consistently outperform a simple 27-day persistence model. Operational physics-based (3D MHD) models, in particular, struggle to accurately forecast high-speed streams associated with co-rotating interaction regions. These regions arise from fast solar wind generated by coronal holes, which are not clearly captured in the magnetogram maps routinely used as inputs. While recent empirical and data-driven methods have shown relatively better performance, significant challenges remain.

Our approach integrates lessons from prior models into what we believe represents the current state-of-the-art. Specifically, we use GONG synoptic maps (magnetograms) and full-disk SDO EUV images as inputs to a neural network. This network estimates the optimal inner boundary condition for the radial solar wind velocity profile at 10 solar radii, which is then propagated to 1 AU using a simplified 1D hydrostatic model.

The key innovation lies in seamlessly integrating the physics-based model within the neural network, creating a true physics-informed ML framework.

We will present validation metrics to assess the model’s performance and discuss plans to make the forecast outputs available to the community 24/7 via the swx-trec.com portal.