Physics-Informed Weather Analytics for Energy Traders

Written by: Olivier Lam, Physical AI Team, Jua.ai AG | Last updated: July 11, 2026

Key Takeaways for Trading Teams

• Physics-informed weather analytics embeds conservation laws directly into forecasting models, so outputs respect physical constraints by design, not by later correction.
• Hybrid AI weather models combine neural network speed with governing equations like Navier-Stokes, which removes the unphysical outputs common in unconstrained machine-learning forecasts.
• EPT-2, Jua’s Earth Physics Transformer, enforces mass, momentum, and energy conservation at the representation level and outperforms ECMWF HRES and ensemble models across multiple variables and lead times.
• Energy traders using Jua for Energy can run live benchmarks against 25+ models and complete backtests in under five minutes, turning accuracy gains into millions in annual savings for wind and solar portfolios.
• Validate physics-constrained forecasts against your current provider and see the accuracy impact in minutes.

The Limits of Today’s Weather Forecasting Stack

Numerical weather prediction (NWP) has led operational forecasting for forty years by decomposing the atmosphere into three-dimensional grid cells and solving differential equations inside each one. A single NWP simulation consumes approximately 8,400 kWh and costs €1,000–€20,000 to run on high-performance computing infrastructure. That compute ceiling caps update frequency at two to four runs per day, which leaves energy traders working with stale numbers between cycles.

Pure data-driven AI models reduce compute cost but introduce a different failure mode. A standard transformer applied naively to atmospheric data can produce outputs that violate conservation laws, and mass-wind field decoupling increases with forecast horizon in black-box AI models, as documented in the HMgNO hybrid framework study published in npj Climate and Atmospheric Science. Aurora’s autoregressive 6-hour rollout compounds errors across up to 84 steps at 21-day leads, which produces a practical predictability horizon that collapses beyond 7–10 days, as measured in independent Aurora verification research. Neither pure NWP nor unconstrained AI alone meets the accuracy, cost, and update-frequency requirements of production energy trading.

These complementary failure modes, NWP’s compute ceiling and unconstrained AI’s physics violations, point to a structural hybrid solution. A model that learns physics from data while enforcing conservation constraints by construction captures the speed of neural inference and removes the hallucination risk of unconstrained architectures.

Compare 25+ models head-to-head on your region in under 5 minutes, without waiting for the next NWP run.

How EPT Embeds Physics Inside Neural Networks

Physics-informed neural network forecasting follows a straightforward principle: the loss function used during training includes terms that penalize violations of physical laws, and the model architecture encodes those constraints into its latent representation. Two implementation strategies exist. The first adds physics-based penalty terms to a standard data-driven loss, so the model can violate conservation laws but receives a penalty when it does. The second embeds the governing equations directly into the forward pass, which keeps outputs physically consistent by construction instead of relying on gradient pressure.

The Navier-Stokes equations govern fluid momentum in the atmosphere, and embedding them as conservation-loss terms forces the model to respect mass continuity and momentum balance at every grid point and every lead time. Models such as ClimODE apply ordinary differential equation solvers to atmospheric dynamics and treat the forecast as a continuous-time integration problem. CREDIT (the Climate Response Emulator for Deep Inference and Training) couples a neural backbone with physical parameterization schemes to preserve energy budgets across the vertical column. PASSAT (Physics-Augmented Spatiotemporal Attention Transformer) uses attention mechanisms conditioned on physical state variables to enforce geostrophic balance in mid-latitude flow.

EPT takes the second, stricter approach. The Earth Physics Transformer is a general spatiotemporal transformer foundation model that learns the governing physics of complex systems directly from observational data, in a latent representation that is integrated forward in time. Conservation of mass, momentum, and energy is enforced at the representation level, not as a post-processing correction and not as a soft penalty, but as a structural property of the forward pass. The architecture cannot produce outputs that violate those laws the way a generic transformer applied naively to physics would. This distinction separates physics-informed machine learning weather, which embeds constraints by construction, from physics-adjacent machine learning weather, which approximates them through regularization.

EPT-2 was trained on more than 5 petabytes of weather and climate data from over 120 distinct sources, including geostationary and polar-orbiting satellites, surface station networks, national radar networks, ocean buoys, ERA5 reanalysis, and operational ECMWF HRES initial-condition fields. Training required 8 × H100 GPUs over 10 days, compared with 32 × A100 GPUs over 18 days for Microsoft Aurora, which uses four times more GPUs and a substantially longer training cycle.

See conservation laws enforced by construction and run EPT against your current provider’s outputs.

Benchmark Results That Matter for P&L

EPT-2 benchmark results appear in the peer-reviewed technical report arXiv:2507.09703, and EPT-1.5 results appear in arXiv:2410.15076. Evaluation uses StationBench, Jua’s open-source benchmarking methodology, verified against more than 10,000 real ground stations with no post-processing or station fine-tuning. Physics-informed weather versus GraphCast and ensemble forecasting physics constraints are addressed directly in the comparison table below.

A 1 GW wind portfolio that gains four percentage points of forecast accuracy saves roughly €1.5 million per year, and that figure scales linearly across multi-GW portfolios. A 1 GW solar portfolio at the same four-percentage-point accuracy gain saves approximately €3 million per year.

Quantify your accuracy gain, using the benchmark that turned skeptics into champions in five minutes.

From Forecast to Trade: Athena and StationBench in Practice

Jua for Energy exposes EPT and Athena through a unified product surface. Athena turns raw physics predictions from EPT-2 into actionable outputs by reading market context and modeling participant behavior. A trader types a natural-language objective such as “backtest a wind-ramp strategy on EPT-2e over the last two winters,” and Athena plans, calls tools, evaluates intermediate outputs, and returns a briefing, a benchmark, a backtest, or a custom widget. Typical queries resolve in approximately 90 seconds, and a full backtest completes in about five minutes. Physics-informed weather analytics for energy trading is the first market Athena supports, and the agent architecture will expand to other physical-economy verticals.

Live model benchmarking on the Jua platform covers more than 25 models, including 10 proprietary AI models from the EPT family plus 15 third-party NWP and AI models such as ECMWF HRES, ECMWF ENS, Microsoft Aurora, GFS GraphCast, and DWD ICON. Traders can select any region, any variable, and any time window, with results in under five minutes. EPT-2 benchmark results are verified through StationBench, Jua’s open-source evaluation methodology, against more than 10,000 real ground stations with no post-processing or station fine-tuning. A 2026 review by Luo et al. identifies rigorous station-based intercomparison frameworks as essential for credible AI weather model evaluation, and StationBench is Jua’s production implementation of that standard. Jua’s models can natively forecast up to a 5 km resolution, and the Jua platform supports up to 1 km resolution for product-level comparisons.

Pipe Jua forecasts into your own models. Install pip install jua to start, or schedule StationBench on your region through the platform.

Frequently Asked Questions

What makes EPT different from other physics-informed neural networks for weather forecasting?

EPT prevents conservation-law violations by design instead of merely penalizing them during training. Unlike most physics-informed neural networks that add conservation-loss terms to a standard objective, EPT’s architecture enforces the mechanism described in the Core Concepts section at the representation level, which makes physically impossible outputs structurally impossible. EPT-2 is also benchmarked externally against more than 10,000 real ground stations via StationBench, with results published in the peer-reviewed technical report arXiv:2507.09703 and no post-processing or station fine-tuning.

Is Jua a weather AI company?

Jua operates as a foundation model and agent company rather than a single-domain weather vendor. EPT functions as a general physics foundation model, and Athena functions as an AI agent. Jua for Energy is the first applied product where both appear together, similar to the relationship between Anthropic and Claude Code as a horizontal AI platform with a flagship vertical product. The atmosphere is the first physical system EPT has been fine-tuned for, and energy trading is the first market Athena supports, with future expansion planned into plasma fusion, aerospace, materials, and other physical-economy domains.

We already subscribe to ECMWF. Why would we add Jua for Energy?

Jua for Energy complements rather than replaces ECMWF. Most customers keep their ECMWF subscription and run Jua for Energy alongside it, and ECMWF AIFS even runs natively on the Jua platform. Jua for Energy displaces the plumbing around the ECMWF feed, including the in-house grib pipeline, the manual benchmarking harness, the morning-briefing analyst, and the spreadsheet stitching. The 7–9 a.m. manual prep routine compresses into a single workspace, refreshed up to 24 times per day, where every model, including ECMWF HRES, ENS, AIFS, Aurora, GraphCast, and the full EPT family, runs under one schema and one API.

How quickly can we validate EPT-2 against our current provider?

The live benchmark runs on the spot and typically completes in under five minutes for a full backtest, as described in the Implementation section. The Jua platform accepts any region, any variable, and any time window. A prospect selects their highest-stakes region and variable, adds their current provider alongside EPT-2, and the platform returns a side-by-side accuracy comparison immediately. Teams that prefer programmatic access can run the same benchmark through the Python SDK.

Can Jua for Energy integrate with our internal trading and risk systems?

Jua integrates into existing trading and risk stacks through a REST API that exposes more than 25 models, with Apache Arrow support for large payloads and a Python SDK installable via pip install jua. Hindcast data is available across multiple Jua and third-party models for backtesting, and ENTSO-E grid data integrates directly for European power-market workflows. Quant teams pipe Jua forecasts into their own systematic models, and utilities and trading houses route them into existing dispatch and risk tools. Integration that often takes a quarter to build elsewhere stands up in days, and full API documentation is available at docs.jua.ai.

Run the five-minute benchmark that closes deals on your region, your variables, and your current provider.