Statistical and AI modeling for weather forecasting frequently relies on the R programming language due to its extensive suite of statistical libraries. As weather forecasting models become more complex and data-intensive, R’s computational limitations become evident, especially for large-scale matrix operations needed for modern weather prediction systems. This paper proposes a hybrid approach that integrates the parallel processing capabilities of GPUs through CUDA while leveraging R’s data processing strengths. The focus is accelerating two primary computational bottlenecks within a physics-informed statistical modeling framework for wildfire aerosol propagation. By implementing these components in CUDA, the study aims to reduce computation time while maintaining the accuracy and functionality of the original R-based methodology. The results demonstrate a substantial reduction in computation time, with 1019x speedup in the transition matrix computation and up to 486x speedup in the Gibbs sampling algorithm, leading to an overall 28x reduction in total processing time. This acceleration facilitates higher-resolution modeling and near real-time tracking of wildfire smoke dispersion, thereby improving environmental monitoring and decision-making capabilities.