N252-105-0130 MICROCAST: A Unified Hybrid Forecasting System for Naval Operations

N252-105 Executive SummarySBIR Topic N252-105 identifies a critical gap between global model predictability and the high-resolution environmental forecasts needed for tactical naval operations—particularly at sub-kilometer horizontal and ~10?m vertical scales in complex, dynamic theaters. To bridge this gap, we propose MICROCAST (Machine Intelligence for Climate Resolution and Operational ForeCAST)—a hybrid, ML-enhanced downscaling system built in the Julia open-source programming language. MICROCAST integrates global spectral modeling, physics-informed machine learning, and dynamically triggered large-eddy simulations (LES) into a unified, adaptive forecasting pipeline designed for real-time use in Navy operations. The core architecture includes:* SpeedyWeather.jl, a global spectral model that captures synoptic-scale dynamics with modular physics and high-performance GPU acceleration;* Universal Differential Equations (UDEs), which embed trainable neural surrogates within differential equations to correct model biases while preserving physical constraints;* Oceananigans.jl, a GPU-accelerated solver that launches LES domains dynamically in mission-relevant regions to resolve coastal wind shifts, turbulence, and boundary-layer effects.The system uses machine learning to trigger high-resolution domains, generate initial/boundary conditions, and optionally adjust model parameters in real time. Built entirely in Julia, MICROCAST is scalable across centralized HPC systems and low-power tactical platforms—enabled by Julia’s Reactant.jl compiler for deployment across CPU, GPU, and TPU environments. Phase I–III Goals:In Phase I, we establish feasibility through validated prototypes and performance comparisons. In Phase II, we will deliver an end-to-end prototype producing real-time, sub-kilometer forecasts, validated in field conditions. In Phase III, we transition the system to operational use with a scalable, globally relocatable design that performs skillfully in forward-deployed and limited-compute environments. Key Personnel:Our team includes Dr. Josh Day (Rallypoint 1, PI), expert Julia data scientist; Dr. Milan Klöwer (Oxford), creator of SpeedyWeather.jl and a leader in ML-integrated climate modeling; and Dr. Chris Rackauckas (JuliaHub/MIT), creator of SciML and global authority on hybrid ML-physics simulation. This team is backed by JuliaHub and supported by collaborative engagement with Naval Postgraduate School researchers.Transition & Commercialization:MICROCAST will transition to Navy operations through integration with existing environmental systems and deployable compute platforms. Commercial applications include agriculture, energy, transportation, and wildfire forecasting—delivered via mobile platforms and integrations with services like Windy.