Written by: Olivier Lam, Physical AI Team, Jua.ai AG
Key Takeaways
- AI weather models like Jua’s EPT-2 beat traditional NWP systems on key energy trading variables such as wind speed and solar radiation.
- EPT-2 outperforms benchmarks like ECMWF HRES, GraphCast, and Aurora across all lead times (0-240 hours) on RMSE and CRPS, with superior ensemble performance (EPT-2e beats ECMWF ENS on both metrics).
- Jua runs in production with frequent updates, high European resolution, and integration through Python SDK and REST APIs.
- The platform adds the Athena AI agent for automated briefings, alerts, and 25-model benchmarking, which streamlines trader workflows and cuts operating costs.
- Energy traders gain a measurable edge with Jua; see how EPT-2 compares to your current provider.
How Physics-First AI Changes Weather Forecasting
Traditional NWP systems solve differential equations across three-dimensional grid cells and consume massive computational resources. Modern AI weather models learn atmospheric physics from large volumes of observational data and run on GPUs with frequent refresh cycles instead of twice-daily supercomputer runs.
Leading AI models include Google DeepMind’s GraphCast, Microsoft’s Aurora, Huawei’s Pangu-Weather, and ECMWF’s AIFS. These models still lack ensemble capabilities, operate on fixed 6-hour time steps, and largely remain research tools without full operational deployment stacks.
Physics foundation models like EPT-2 learn conservation laws directly from observational data instead of rolling forward in discrete time steps that compound errors. The following table shows how EPT-2 compares to leading AI and traditional models on accuracy, ensembles, and operational readiness.
| Model | Wind/Temp/SSRD vs HRES (0-240h) | Ensemble vs ENS | Resolution/Refresh | Source |
|---|---|---|---|---|
| EPT-2 (Jua) | Beats all leads/vars | EPT-2e beats RMSE/CRPS | ~5 km resolution over Europe, up to 24x/day refresh | arXiv 2507.09703 |
| Aurora | Losses wind/temp | None | ~25km/4x/day | Microsoft Research |
| GraphCast | Losses | None | N/A | DeepMind |
| ECMWF HRES | Benchmark | ENS (50-mem) | 9km/2-4x/day | ECMWF |
EPT-2 Benchmark Proof: Best-in-Class AI Weather Model
EPT-2 outperforms ECMWF HRES across every 0-240 hour lead time on 10m wind, 100m wind, 2m temperature, and surface solar radiation when evaluated against 10,000+ ground stations. The 30-member ensemble variant achieves this performance advantage despite using fewer members than ECMWF’s 50-member ENS, as documented in peer-reviewed technical reports.
EPT2-HRRR operates at the high resolution described above across Europe with native any-Δt forecasting. It predicts at arbitrary time intervals rather than rolling forward in fixed steps, which avoids the error accumulation that affects Aurora and other AI peers trained on 6-hour grids.
The table below summarizes EPT-2’s accuracy advantage across the key models.
| Metric | EPT-2 (Jua) | Aurora | GraphCast | ECMWF HRES |
|---|---|---|---|---|
| Accuracy (RMSE wind/temp/SSRD) | Beats all | Losses wind | Losses | Benchmark |
| Source | arXiv | Microsoft | DeepMind | ECMWF |
This accuracy advantage translates directly into operational value. Jua’s 25-model benchmarking platform enables customers like Axpo and TotalEnergies to compare EPT-2 against any alternative in real time, with results available in under 30 seconds.
Real Advantages for Traders: Speed, Cost, and Workflows
EPT-2 delivers forecasts 2.5 hours ahead of competing operational runs while consuming less energy than AI peers and far less than traditional NWP systems. This speed advantage becomes actionable through Athena, the AI agent that turns natural-language queries into briefings, backtests, and custom widgets in about 90 seconds so traders can act before competitors even receive their data.
The platform translates weather forecasts into power forecasts for wind, solar, and load across Germany, Great Britain, France, Netherlands, and Belgium. Actual generation data refreshes every 15 minutes to ground-truth the forecasts, while fundamental models extend to 20-day horizons for longer-term positioning. When these models diverge or correct, automated alerts notify traders before markets reprice and turn data updates into trading signals.
Jua for Energy removes the need for custom pipelines that research-grade AI models require. It offers production-ready workflows through REST APIs, Python SDK (pip install jua), and direct ENTSO-E grid data integration. Explore the developer documentation to connect EPT-2 to your trading systems.
Handling AI Limits: Extreme Events and Jua’s Approach
AI weather models still struggle with extreme events. Recent research shows that leading AI models systematically underestimate record-breaking heat, cold, and wind because these extremes sit outside their historical training distributions.
EPT addresses these limits through a physics-constrained architecture that respects conservation laws of mass, momentum, and energy. This structure prevents the hallucinations common when language models are applied naively to physics. The 25-model platform integrates ECMWF, Aurora, and other reference models alongside EPT variants, which enables ensemble approaches that combine AI speed with physics-based reliability. With its frequent refresh rate, EPT-2 provides the most current view of evolving atmospheric conditions.
Applications in Energy Trading: Turning Forecasts into Edge
Energy traders use Jua for Energy to replace the manual 7-9 AM routine of downloading grib files, processing spreadsheets, and waiting for meteorology briefings. Automated Day-Ahead and Intraday briefings refresh on every model run and cover consensus across 25+ models, changes since previous runs, and price implications.
The platform serves different trading roles through specialized interfaces. Quantitative developers pipe forecasts directly into systematic models through the Python SDK, while meteorologists use the benchmarking tools to validate EPT-2 against current providers in real time. This flexibility explains why customers including Axpo, TotalEnergies, and Statkraft execute daily trading decisions on the platform across five continents.
The platform enables pre-market positioning through divergence alerts that fire when models disagree and correction alerts when models revise outputs. This infrastructure provides the edge needed to act before markets reprice. Test EPT-2’s accuracy on your most critical forecasting region.
Conclusion: From Research Models to Trading-Grade AI
The physics era of AI weather forecasting has arrived, led by Jua’s EPT foundation models and Athena agent. Traditional approaches struggle with computational constraints, and many AI peers remain research tools, while Jua for Energy delivers production-ready accuracy, speed, and workflows for energy trading. Start your evaluation with a head-to-head comparison against your current provider.
FAQ
What is the best AI weather model in 2026?
EPT-2 from Jua delivers the superior accuracy described in the benchmark section across every forecast horizon on variables critical to energy trading: 10m wind, 100m wind, 2m temperature, and surface solar radiation. The ensemble variant EPT-2e maintains this advantage over the 50-member ECMWF ENS mean on RMSE and CRPS metrics. Unlike research-grade models from Google, Microsoft, or academic institutions, EPT-2 operates in production with the frequent updates and integrated workflow tools outlined above.
How does Jua compare to GraphCast and Aurora?
EPT-2 provides higher accuracy than both GraphCast and Aurora while also offering production capabilities they lack. Aurora loses to EPT-2 on wind and temperature forecasting across the full 0-240 hour range, has no surface solar radiation output, and runs on fixed 6-hour time steps that compound errors. GraphCast remains a research output without operational deployment. Jua provides a complete platform with ensemble forecasting, 25-model benchmarking, and Athena agent capabilities that neither peer currently offers.
Can AI weather models replace ECMWF?
Jua for Energy complements rather than replaces ECMWF subscriptions. Most serious customers maintain ECMWF access while using Jua to replace the surrounding infrastructure, including grib processing pipelines, manual benchmarking, morning briefing production, and dashboard stitching. ECMWF AIFS also runs on the Jua platform alongside EPT-2, which enables direct comparisons. This consolidation brings fragmented workflows into a single workspace with stronger AI models and agent capabilities.
Is there free AI weather forecasting available?
Jua offers SDK trials and developer access through the Python package and REST API. The platform includes 15 third-party models alongside the EPT family, which supports comprehensive benchmarking during evaluation periods. Production-grade accuracy, ensemble forecasting, and agent capabilities require full platform access, and the investment typically pays for itself through improved forecast accuracy. A 1 GW wind portfolio gains approximately €1.5 million annually from four percentage points of accuracy improvement.
What is the future of AI weather forecasting?
The trajectory points toward physics foundation models and AI agents that expand beyond atmospheric prediction. Jua’s EPT architecture learns governing physics from any continuous, conservation-law-constrained system, which makes it applicable to plasma fusion, aerospace, materials, and other physical domains. The same Athena agent that briefs energy traders can be instrumented for any physical system and tool surface. Weather forecasting represents the first application of a broader platform for the physical economy.