Presented at the Naval Applications of Machine Learning (NAML) Conference, February 2025
Extreme and unpredictable weather conditions are worsening existing security risks. Existing numerical weather prediction (NWP) models are computationally intensive and expensive, requiring long processing times and significant compute resources. Foundational ML based methods require less time and less computationally expensive resources.
We introduce a novel method using ML for processing climate regions by utilizing cylindrical layers within a Vision Transformer neural network. By replacing NWP, equation-heavy ensemble calculations, with cylindrical convolutional tokenization, we dramatically reduce computational overhead without sacrificing precision and efficiency. This state-of-the-art solution delivers high-resolution, adaptable time-series analysis across multiple data modalities, unlocking faster insights, lower costs, and greater flexibility for organizations facing complex, data-driven challenges.
This architecture enables improved accuracy for long-term horizon weather forecasting, enhancing operational planning capabilities. Its adaptability opens the door to forecasting challenges in other domains such as disease outbreaks and wildfires—where timely, spatially-aware predictions are critical.