Ocean, Supercomputers and AI in action!
AI models engineered on CINECA HPC infrastructure are a core element of the services that Gmatics developed as part of the DAMAS – Data-driven Model for the Analysis of Sea-state experiment. DAMAS successfully reduced the execution time and increased the resolution and the accuracy of sea circulation models and height of waves prediction, providing as a result:
- AI models that replicate the ENEA MITO and WAVE physical forecast models and produce forecasts over the next 5 days with an hourly temporal resolution;
- the nowcast chains for wave and circulation for the next 12 hours,
- hindcast pipelines for the historical analysis of wave and circulation sea parameters and
- a graphical user environment
122.586 core hours have been consumed and 44 TB of data have been processed in order to reduce the AI model runtime for the whole Mediterranean Sea to 15 minutes (the experiment’s final target was 30 minutes), increase the geospatial resolution to 1×1 Km, achieve a point accuracy with a Root Mean Square Error between DAMAS and ENEA results below 5 % for each variable (total peak and mean period, total mean wave direction, sea surface wave significant height, sea surface height, surface salinity, surface temperature, surface wind u-component and v-component speed), and increase the areal accuracy above 90% for each variable.
DAMAS significantly increased the resolution and the accuracy of sea circulation models and height of waves prediction, also reducing their execution time
These results were achieved during the 9 months duration of the experiment by first matching the performance of the ENEA model and then improving the geospatial and temporal resolution, which required the fine-tuning of the AI model using satellite data, ensemble techniques and transfer learning techniques.