A newly demonstrated neural-network surrogate workflow substantially accelerates the evaluation of empirical geophysical models while maintaining sub-percent mean approximation errors across tested domains. In evaluations of multiple Earth and space science models, the surrogate architectures successfully reproduced complex reference outputs, with the worst reported scalar or component mean absolute percentage error limited to just 0.981 percent.
The research was conducted by Volodymyr Kashyn and Vasyl Choliy and published in the journal Earth Science Informatics. Their study addresses the significant computational demands associated with traditional empirical models by establishing a standardized machine learning pipeline designed to replace expensive calculations with rapid neural-network inference without sacrificing physical precision.
To test the methodology across diverse physical phenomena, the authors evaluated four standard reference models: the Naval Research Laboratory Mass Spectrometer and Incoherent Scatter Radar Extended model (NRLMSISE-00), the International Reference Ionosphere 2020 (IRI 2020), the International Geomagnetic Reference Field 14 (IGRF-14), and the Earth Gravitational Model 1996 (EGM96). These established empirical formulations represent atmospheric, ionospheric, geomagnetic, and geopotential systems frequently utilized in Earth and space science applications.
The proposed workflow provides an end-to-end framework governing every phase of surrogate development. It encompasses input-domain sampling, spatial and temporal representation, input and output normalization, neural-network architecture selection, model training, and the restoration of physical outputs from normalized values. The protocol concludes with a rigorous dual evaluation focusing on both approximation accuracy across the input domains and computational performance improvements relative to the baseline code.
Computational benchmarking showed marked speedups when restricting both the original reference C++ models and neural-network inference to a single central processing unit (CPU) thread on an AMD Ryzen 5900X processor. When processing large input batches containing up to two million samples, single-threaded surrogate execution achieved speedups of 4.3× for the NRLMSISE model, 332× for the IRI model, 126× for the IGRF-14 magnetic-field model, and 7.08× for the EGM96 geopotential model relative to the reference implementations.
Hardware acceleration using a dedicated graphics processing unit (GPU) yielded even more substantial performance increases when measured against the same single-threaded C++ baselines. Deploying the neural-network surrogates on an NVIDIA RTX 4070 delivered speedup factors of 39× for NRLMSISE, 921× for the geomagnetic field model, and 98.8× for the geopotential model. The highest computational gain was recorded for the International Reference Ionosphere model, where GPU inference achieved an acceleration factor of 18 188×.
Overall, the results demonstrate that the workflow can reliably prepare neural-network surrogates that reproduce reference-model outputs with sub-percent mean errors and greatly reduced execution times. However, the investigators noted an important scope limitation in their current findings. While individual model evaluations demonstrated substantial acceleration, the effect of integrating these surrogate models into complete orbit-propagation, orbit-determination, and observation-modelling workflows remains to be evaluated in operational settings.
Subject of Research: Neural-network surrogate workflow for empirical geophysical models
Article Title: A neural-network surrogate workflow for accelerating the evaluation of empirical geophysical models
Article References: Kashyn, V., & Choliy, V. (2026). A neural-network surrogate workflow for accelerating the evaluation of empirical geophysical models. Earth Science Informatics, 19(12), Article 250. https://doi.org/10.1007/s12145-026-02308-2
Image Credits: AI Generated
DOI: 10.1007/s12145-026-02308-2
Keywords: neural networks, surrogate models, geophysical models, NRLMSISE-00, IRI 2020, IGRF-14, EGM96, computational performance, approximation accuracy, GPU inference, CPU benchmarking, AMD Ryzen 5900X, NVIDIA RTX 4070, orbit propagation, orbit determination, Earth science informatics, observation modelling
News Source: Cassandra Pierce. (November 16, 2026). Neural Surrogates Accelerate Geophysical Models With Sub-Percent Mean Error. Scienmag.



