A Neural-Network Surrogate Workflow for Accelerating the Evaluation of Empirical Geophysical Models

<title>Abstract</title> <p>This paper presents a neural-network surrogate workflow for accelerating the evaluation of empirical geophysical models used in Earth and space science applications. Four reference models are considered: NRLMSISE-00, IRI 2020, an IGRF-14-based magnetic-field model, and an EGM-96-based geopotential model. For each model, a surrogate was trained on data generated by the corresponding reference C + + implementation. The workflow covers input-domain sampling, spatial and temporal representation, input and output normalization, architecture selection, training, restoration of physical outputs, and evaluation of approximation accuracy and computational performance. Within the tested input domains, all surrogates achieved sub-percent mean approximation errors for the considered outputs, with the worst reported scalar/component mean absolute percentage error equal to 0.981%. They also substantially reduced execution time relative to the reference implementations. When both the reference C + + models and neural-network inference were restricted to a single CPU thread on the same AMD Ryzen 5900X, the best speedups for input batches of up to two million samples were 4.3× for NRLMSISE, 332× for IRI, 126× for the magnetic-field model, and 7.08× for the geopotential model. Relative to the same single-threaded C + + baselines, GPU inference on an NVIDIA RTX 4070 achieved speedups of 39×, 18188×, 921×, and 98.8×, respectively. These results show that neural-network surrogates can preserve sub-percent mean approximation accuracy while substantially reducing the cost of repeated geophysical-model evaluation. The workflow provides an efficient data-driven approach for large-scale Earth and space science applications, including satellite orbit propagation, observation modelling, and simulation.</p>