Today, we're publicly introducing the EPT-2 family, the latest advancement in our Earth Physics Transformer (EPT) series. This release represents meaningful progress in our mission to simulate the future by learning our universe's physics directly from data.
Why it Matters
Energy is the foundation of modern civilization. Every aspect of human progress, from the devices we use to the cities we build, depends on reliable, abundant power. We're in the middle of the largest energy transition in human history. We are electrifying transportation, heating our homes with heat pumps, powering increasingly energy-hungry AI systems, and transitioning to renewables that are essentially free but fundamentally volatile.
This creates an unprecedented forecasting challenge. Traditional power systems were predictable: you turned on a coal plant when you needed electricity. Now, our energy supply depends on when the wind blows and the sun shines, while demand spikes unpredictably as millions of EVs charge and AI datacenters come online. Getting this wrong doesn't just mean higher electricity bills; it means blackouts, economic disruption, and the potential failure of our energy transition.
Beyond energy, the ripple effects touch everything: food security depends on predictable weather for crops, global supply chains need to be resilient, and trillions in climate-sensitive infrastructure hangs in the balance. Yet our forecasting tools remain fundamentally unchanged from decades ago.
We've always believed there's a better way. And we've spent the past three years building it.
Our bet on AI for earth simulation
Two years ago, on March 1st, 2023, we deployed Vilhelm, the world's first operational AI weather forecasting system. It was the first time a global deep learning model produced native hourly forecasts in production. It was also the first time a weather model produced 1x1 km precipitation forecasts at global scale. That moment marked a turning point, not only for Jua but for how the world might think about modeling Earth systems.
At the time, the broader meteorological community was still exploring whether AI models could outperform traditional numerical weather prediction. Today, that path has proven successful, and leading organizations are increasingly embracing AI-based approaches alongside their existing capabilities.
Since our first release, we've continued to push forward: releasing EPT-1, then EPT-1.5, and now EPT-2. Each iteration brought deeper physical fidelity, more flexibility, and greater performance.
The EPT-2 Family
Fluid physics are incredibly hard to simulate, especially at the scale of Earth's atmosphere. A single change in one spot can ripple across continents within days. Traditional weather models break the Earth into boxes and solve equations for each box separately, missing the deep interconnections that drive weather patterns. For the first time, EPT-2 treats Earth as the interconnected system it truly is.
The EPT-2 family consists of three specialized models:
- EPT-2: Our deterministic model for precise weather forecasting
- EPT-2 Early: Our early-availability model for time-critical applications
- EPT-2e: Our ensemble model for probabilistic weather forecasting, especially for the long term
EPT-2e builds on the deterministic core of EPT-2 and sets a new state of the art in global weather prediction.
In benchmarks against the most advanced models available, including Microsoft's Aurora and ECMWF's ensemble mean, EPT-2e achieves record-breaking accuracy. It delivers better forecasts across every key variable and lead time, and does so more efficiently.
Here's what sets it apart:
Global interconnectedness: Unlike traditional models that solve equations for isolated grid boxes, EPT-2e captures the complex fluid dynamics that make Earth's atmosphere a single, interconnected system.
Superior accuracy across all horizons: EPT-2e delivers significantly lower forecast errors on wind, temperature, and solar radiation from short-term to long-term predictions. It also forecasts further into the future than any existing model.
Operational speed advantage: It produces global, high-resolution forecasts faster and earlier than existing models. In weather-sensitive industries, this timing unlocks critical advantages. In power trading, it means a significant edge.
Native flexibility: Unlike models trained or executed on a fixed temporal resolution, EPT-2e can predict any timestep of choice.
Trust Comes from Performance
We benchmark EPT-2e using a multi-layered evaluation strategy. Gridded initial condition comparisons and real-world weather station data (via our WeatherReal dataset) are standard in meteorology, but few AI models hold up across both. EPT-2e does. It outperforms every baseline we've tested, across all lead times.
The full technical report is public with a link at the end of this page. Our benchmarking library is open source and available on GitHub.
Early Access and Deployment
Over the last few months, we quietly invited a group of existing customers and new partners (utilities, energy traders, operators) into the next phase of our platform. They've been using and evaluating the EPT-2 series internally. The feedback has been clear and enthusiastic.
The EPT-2 family is now available on the Jua Earth Intelligence Platform. For the first time, a globally consistent, high-resolution, AI-native forecast suite (including deterministic, early, and ensemble models) is just an API call away.
Getting Started:
- Graphical Platform: Access EPT-2e directly through our web interface on our Earth Intelligence Platform
- API Access: Grab an API key from the platform and integrate EPT-2e into your applications using our Python SDK:
pip install jua
Where We're Headed
EPT-2e is not a finished product. It's a capability layer that we'll continue to extend beyond weather into broader simulation of our physical surrounding: energy production & extraction, hydrology, vegetation, and beyond into broader fluid mechanics & inverse problems. Our long-term vision is to unlock the ability to simulate the future itself. By learning our physical world directly from data, we're building toward a world where every critical decision can be informed by precise predictions of what will happen next.
Our philosophy is simple: everything we can simulate, we should simulate. Reality is fragile and expensive. Every test in the real world carries risk: every power trade made without knowing tomorrow's wind patterns, every decision to shut down wind parks without understanding the precise timing of weather fronts, every supply chain routed without predicting disruptions. Even beyond forecasting, every wind blade not designed optimally represents millions in lost efficiency that could have been avoided through better and faster simulation. Simulation lets us explore infinite possibilities without the cost of failure.
The economic value of this capability is extraordinary. Industries managing trillions in assets will fundamentally transform how they operate. This release brings us another step closer to that future.
A Final Word
To the team: thank you. Your focus, rigor, and speed over the past year have been extraordinary. To our partners and investors: your belief in this mission made this possible.
Explore EPT-2e. Build with it. Push its limits. Let's make the Earth more understandable together.
For detailed technical information about the EPT-2 series, benchmarks, and performance metrics, check out our comprehensive technical report.
Marvin Gabler, CEO