Solar Wind Prediction AI Tournament
Operational Machine Learning Competition for 72-Hour Solar Wind Forecasting
Introduction
Can machine learning (ML) systems predict solar wind speed three days ahead under the same real-time constraints faced by operational forecasting systems?
The Frontier Development Lab (FDL.ai) Benchlab is a competition and benchmarking environment designed to test scientific ML systems under realistic deployment constraints: automated execution, standardized containers, runtime limits, and evaluation against future observations.
Solar wind forecasting is central to space weather preparedness, influencing everything from satellites to power grids, yet reliable multi-day prediction remains one of heliophysics' greatest open problems. Solar wind speed is a core operational variable because it captures the arrival and evolution of different solar wind regimes, including slow wind, fast streams, and stream interaction regions that can drive geomagnetic activity on Earth. It is also continuously measured near Earth, making it suitable for real-time evaluation, while still remaining difficult to forecast accurately several days in advance.
This challenge asks whether data-driven systems can improve forecast skill while remaining reproducible, robust, and operationally deployable. FDL and the University of Colorado Boulder (SWx-TREC) invite data scientists and heliophysicists worldwide to join FDL Benchlab and bridge the gap between offline model development and true forward-in-time, real-time deployment.
Pre-register to receive launch materials, technical documentation, templates, and briefing information ahead of launch on July 31, 2026.
The Challenge
Beyond Static Data: A True Operational Loop
Unlike traditional machine learning competitions, which rely on static, historical datasets, this tournament tests workflows inside a live, stateless, and constraint-driven operational loop using unseen future data.
Detailed target variables, forecast formatting requirements, and evaluation APIs will be released at competition launch.
Atomic Inference
Models must be entirely stateless and self-contained, utilizing only data available at the exact inference timestamp and automatically fetching their own dependencies.
Continuous Run Phase
Once participant workflows are frozen, they will run continuously for a minimum of three months (the organizers may extend the evaluation window to capture scientifically relevant solar wind conditions, with extension criteria communicated in advance). Models trigger every 6 hours to generate a 72-hour forecast at a 1-hour resolution.
Zero Leakage Evaluation
Submissions execute on a CPU-based hardware harness and are evaluated against real-time ACE data in the future, i.e., true forecasts that eliminate any information leakage.
Important Dates & Milestones
June – July 2026 | Pre-Registration Open Secure one of the limited spots
July 31, 2026 | Official Launch & Briefing Documentation, templates, and environments released
October 1–10, 2026 | Mandatory Pre-Deploy Week Run pipeline tests via a single command
October 12, 2026 | Submission Close & Code Freeze Final deadline for all workflows
October 13, 2026 | Live Evaluation Run Phase Opens Models run live every 6 hours for the full evaluation period (minimum 3 months)
TBD | Awards Ceremony TBD
Why Benchlab?
Unlike traditional ML competitions
While traditional ML competitions evaluate against static historical datasets, Benchlab evaluates models inside a live operational loop using unseen future data.
Submissions must:
run autonomously
operate without state persistence in the submission environment
tolerate real-world latency and missing inputs
generate forecasts continuously for an extended period