The era of squinting at confusing 10-day weather forecasts on our devices could soon be replaced by a crystal-clear vision of long-range weather predictions, thanks to efforts from the Department of Energy's Oak Ridge National Laboratory (ORNL). Utilizing the formidable capabilities of the world's fastest supercomputer, Frontier, researchers have developed an advanced artificial intelligence model dubbed ORBIT that stands to revolutionize how we receive and use weather information. As reported on ORNL's website, this model is trained to provide highly accurate weather predictions anywhere from days to up to a month in advance.
In a move that could upend everything from daily commutes to emergency response strategies, the 2-exaflop HPE Cray EX supercomputing system of Frontier has trained what is termed the largest AI model for weather forecasting to date—covering our globe down to individual street addresses. Prasanna Balaprakash, ORNL's director of AI programs, explained the significance of this leap in forecasting technology, stating, "These would be forecasts not just for the whole country or state, not just for the local region but for your or my address." According to an interview obtained by ORNL, the goal is to provide people with "hyperlocal, hyperaccurate forecasts" that were previously beyond reach due to the complex nature of climate prediction.
The sophisticated ORBIT system, which stands for Oak Ridge Base Foundation Model for Earth System Predictability, operates with 113 billion model parameters and boasts a stunning 90 to 95% accuracy rate for short-term forecasts. The longer-term forecasts also show promise with accuracy rates ranging between 60 to 80%. This breakthrough earned the ORNL team a finalist nomination for the prestigious Gordon Bell Prize for Climate Modeling. The distinction is notable for recognizing significant advances in high-performance computing used in climate modeling applications.
Importantly, while popular AI models used for various applications, such as OpenAI's ChatGPT, are built upon large language models, ORBIT differentiates itself by relying on a foundation model that intakes an extensive array of weather-specific data, such as cloud formations and temperature fluctuations. During the training process, the involvement of more than 6,000 nodes of Frontier's architecture facilitated the simulations to reach unprecedented speeds, validating ORBIT's capability for dealing with vast amounts of weather data and pattern recognition. Xiao Wang, an ORNL computer scientist, also highlighted ORBIT's design to function "across various computing platforms without being tied to specific hardware" in a statement released by ORNL.
Looking ahead, ORBIT is not only a scientific boon but promises significant energy and cost savings as well. Once the AI model's training is complete, it won't require the same computational heft for updates or usage. As Balaprakash pointed out, "The model only has to be trained once." He anticipates that the same modeling power could eventually be accessed from devices as commonplace as smartphones. This progress aligns with the broader mission of machine learning in meteorology to act as a cheaper, faster, and more efficient surrogate for traditional climate simulations.
Furthermore, integral to ORBIT's reliability will be its capacity for uncertainty quantification, tasking the model with self-assessing the confidence levels in its own forecasts. Dan Lu, an ORNL computational scientist, underlined the importance of establishing trustworthiness by having the model present its predictions with corresponding certainty percentages. Agencies such as the Department of Defense and institutions like the Tennessee Valley Authority are reportedly already showing interest in utilizing ORBIT for their operations, signaling broad implications for strategic planning and resource management linked to weather-dependent activities.
The application of ORBIT is seen as "a starting point," suggesting a continuous journey of refinement and enhancement lies ahead. With potential applications ranging from personal planning to national defense, ORBIT's advancement signals an impending shift in our interaction with the elements, priming society for a future where weather forecasts are as reliable as they are pivotal.