Operational inference of AI-based weather models at FMI

The integration of AI-based models into operational weather forecasting requires not only advances in model development, but also robust, scalable, and reproducible production environments. At the Finnish Meteorological Institute (FMI), we have established an operational framework for running AI-driven forecasting systems using in-house GPU resources, combined with modern software engineering practices and cloud-native orchestration.  

This presentation focuses on the operational inference pipeline of AI-based Aila model, developed within the Anemoi framework as part of the ECMWF Machine Learning Pilot Project, in collaboration with ECMWF and member states. Inference is executed on dedicated GPU infrastructure, enabling both deterministic and ensemble forecasts at high spatial and temporal resolution. We describe the computational requirements of running these models in practice, including GPU allocation strategies, resource utilization, and runtime characteristics, as well as trade-offs between resolution, ensemble size, and latency in an operational setting.  

The inference pipeline runs in a containerized environment, with ecFlow handling scheduling and the jobs themselves are executed on OpenShift. This enables scalable, fault-tolerant deployment that integrates with FMI's production systems and is easily reproducible across platforms. Model execution is coupled with data ingestion, post-processing and model verification, forming an end-to-end workflow suitable for real-time forecasting. 

Supportive tools will also be addressed, along with how the model and code lifecycle is managed. MLflow is used to handle model versioning, configurations, and metadata, ensuring traceability and reproducibility across operational runs. In addition, deployment workflows and operational configurations, such as YAML files defining inference setups, are maintained through GitHub-based version control and continuous integration practices. 

We also discuss practical challenges encountered in operational AI inference, including dependency management, consistency between development and production environments, inconsistency between training and operation environment, and monitoring of system performance.