In this work we investigate accelerating geophysical models using machine learning, by converting traditional empirical models into neural network implementations. We focus on two widely used models: atmospheric, NRLMSISE, which predicts atmospheric temperatures and densities, and ionospheric, IRI, which provides global estimates of ionospheric parameters: electron density, ion composition, temperatures and total electron content (TEC). NRLMSISE and IRI are valuable parts of satellite observation processing software like Kyiv Geodynamics (Juliette). These models are also essential for assessing atmospheric impacts on satellite orbits and laser beam propagation for Earth-surface distance measurements; however, they can be computationally intensive when millions of points are required, and this often limits their broader application. To address this, we reformulated the classic models into neural network architectures. Our methodology involves training the networks using datasets derived from the original models. The resulting NRLMSISE and IRI neural networks reproduce scalar output parameters with mean relative errors below 1%; for IRI ion composition, the mean total variation distance is below 0.25% compared with the reference C++ implementations based on the original algorithms. Execution time benchmarks for NRLMSISE indicate that our approach can achieve a minimum threefold improvement in the same CPU one-threaded environment and at least a tenfold acceleration on GPU. For IRI we got ~300 times speed boost in the same CPU one-threaded environment and from 300 to 18,000 times boost on GPU CUDA depending on input size. In conclusion, the conversion of geophysical models into neural network frameworks not only offers considerable performance gains but also opens the door for wider application of these models in real-time and resource-constrained scenarios.