Structure-Aware Reformulation of Physics-Informed Bayesian Inference for Wildfire Aerosol Forecasting on an Embedded GPU

The GOES-16 and GOES-17 satellites retrieve AOD every five minutes, so a smoke forecasting system must complete each cycle before the next observation arrives. We previously accelerated a Physics-Informed Bayesian State-Space model using datacenter GPUs to meet this requirement. However, datacenter hardware is impractical for remote field deployments, which require compact, low-power embedded modules. We therefore ported the GPU implementation to the NVIDIA Jetson Orin Nano, where the initial implementation took 256 s to execute, leaving little operational margin within the 300 s cycle. A fully single-precision implementation failed because the covariance matrices lost positive definiteness, so double precision was retained. Therefore, additional performance gains must come from optimizing the underlying computations. The baseline implementation executes the statistical recursions directly, resulting in significant computational overhead from not exploiting the available mathematical structure. We optimize these operations by exploiting matrix sparsity, symmetry, and reusable intermediate terms, reducing the Jetson runtime from 256 to 110 s and achieving a 2.33× speedup. All modifications are exact mathematical reformulations rather than approximations, so the reported inferential and forecast quantities agree at the precision at which they are reported, with small floating-point differences remaining in the underlying state arrays. This improvement provides additional headroom for robust edge-based smoke forecasting.