📄 earth_science

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

This paper introduces a neural-network surrogate workflow that significantly accelerates the evaluation of four empirical geophysical models while maintaining sub-percent approximation accuracy, offering a highly efficient data-driven solution for large-scale Earth and space science applications.

Original authors: Volodymyr Kashyn, Vasyl Choliy

Published 2026-08-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Volodymyr Kashyn, Vasyl Choliy

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Cosmic Speedometer: Why We Need to Predict Space Weather Faster

Imagine trying to navigate a spaceship through a stormy ocean, but the ocean isn't made of water—it's made of invisible, shifting gases, magnetic fields, and electrically charged particles that surround our planet. This is the reality for satellites and astronauts. To keep them safe and on course, scientists use complex mathematical recipes called "empirical geophysical models." These models act like a cosmic weather forecast, predicting everything from the density of the upper atmosphere (which can drag a satellite down) to the strength of Earth's magnetic field (which guides navigation).

However, there's a catch. These recipes are incredibly detailed, but they are also slow. Think of them as a master chef who can cook the perfect meal but takes an hour to chop a single onion. When you need to simulate a satellite's path for years, or track thousands of objects at once, waiting for the "chef" to chop every single onion becomes a bottleneck. The computer gets stuck waiting for the math to finish, slowing down the whole mission. This is where the question of speed becomes critical: How do we get the same perfect meal, but in a fraction of a second?

The "Workaround" for Space Math

This paper introduces a clever workaround: a "neural-network surrogate." If the original geophysical models are the master chefs taking hours to cook, these surrogates are like a super-fast food truck that has memorized the chef's recipes. Instead of calculating the complex physics from scratch every time, the neural network is a smart computer program that has been "trained" by watching the master chef cook millions of times. It learns the patterns and relationships so well that it can guess the result almost instantly, with almost the same accuracy.

The researchers tested this idea on four different "chefs" (models) that handle different parts of space weather:

NRLMSISE-00: Predicts the density of the atmosphere (the "air" in space).
IRI 2020: Predicts the ionosphere (a layer of charged particles that affects radio signals).
IGRF-14: Maps Earth's magnetic field.
EGM-96: Maps Earth's gravity and how it pulls on satellites.

The team didn't just guess; they built a rigorous workflow. They fed the original, slow models millions of different scenarios (different times, locations, and solar activity levels) to create a massive dataset. Then, they trained neural networks to mimic the results. The goal was simple: Can the fast "food truck" serve the same meal as the slow "master chef" without anyone noticing the difference?

The Results: A Speed Demon with a Perfect Memory

The answer is a resounding yes, and the speed difference is mind-blowing. The researchers found that these neural networks could predict the results with sub-percent accuracy, meaning the error was less than 1% in almost every case. For example, the worst error they saw for a specific measurement was just 0.981%. In the world of space science, that is practically perfect.

But the real magic is the speed. When the researchers ran the original, slow models on a single computer processor thread, they took a certain amount of time. When they switched to the neural network surrogates on the same single thread, the results came back much faster.

For the NRLMSISE atmosphere model, the neural network was 4.3 times faster.
For the IGRF magnetic field model, it was 126 times faster.
For the EGM gravity model, it was 7.08 times faster.
The biggest winner was the IRI ionosphere model, which was a massive 332 times faster on a single thread.

The results got even wilder when they used a powerful graphics card (GPU), the kind of chip usually found in gaming computers, to run the neural networks. Because these chips are designed to do many calculations at once, the speedup was astronomical. The IRI model became 18,188 times faster on the GPU compared to the original slow model. That's like turning a 5-hour movie into a 1-second clip.

Why This Matters

The paper shows that we don't have to choose between accuracy and speed. We can have both. By using these "surrogate" models, scientists can run simulations that used to take days in a matter of minutes or seconds. This is a game-changer for things like predicting satellite orbits, planning space missions, and understanding how space weather affects our technology on Earth.

The authors are careful to note that these fast models are only as good as the data they were trained on. They work perfectly within the "training zone" (the specific ranges of time and space they saw during training), but if you ask them to predict something totally outside that range, they might get confused. However, for the vast majority of real-world space operations, these neural networks offer a way to navigate the cosmic ocean with a map that is both incredibly detailed and incredibly fast. It's a new tool that lets us look at the universe with sharper eyes and move through it much quicker.

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

Problem Statement
In Earth and space science applications, such as satellite orbit propagation and observation modeling (specifically within the Juliette software package), empirical geophysical models are repeatedly evaluated to describe the near-Earth environment. These models include representations of the neutral atmosphere (NRLMSISE-00), the ionosphere (IRI 2020), the geomagnetic field (IGRF-14), and the geopotential (EGM-96). While the numerical integration of satellite trajectories is computationally efficient, the repeated evaluation of these environmental models creates a significant performance bottleneck. Existing implementations, typically written in C++ and optimized for sequential CPU execution, involve complex control flows and conditional branches that are difficult to port efficiently to data-parallel hardware like GPUs. This limitation hinders large-scale batch computations and dense sampling required for modern orbit determination and simulation workflows.

Methodology
The authors propose a structured workflow to convert these reference C++ implementations into accurate, computationally efficient neural-network (NN) surrogates. The workflow encompasses the following stages:

Dataset Generation and Sampling: Datasets are generated by systematically sampling the input domains of the reference models rather than relying on observational data. The sampling strategy addresses specific challenges:
- Temporal: Uniform distribution over time parameters to ensure unbiased coverage.
- Spatial: To avoid pole concentration inherent in standard latitude/longitude sampling, the authors employ a specific sampling strategy where latitude is transformed via an arcsine function and altitude is sampled based on the cumulative distribution of the spherical shell volume ( $r^3$ ).
- Indices: Solar and geomagnetic indices are either sampled uniformly (to create a general approximation of the algorithm) or derived from historical records (to learn real-world dependencies).
Data Representation and Normalization:
- Inputs: Cyclic time features (day of year, seconds of day) and longitude are encoded using sine-cosine transformations to handle discontinuities. Geodetic coordinates are represented either as sine-cosine pairs (for NRLMSISE) or as a Cartesian unit vector plus altitude (for IRI, IGRF, and EGM) to mitigate coordinate singularities.
- Normalization: Inputs and outputs undergo specific normalization. Features with known fixed ranges use min-max scaling; others use z-score standardization. Logarithmic transformations are applied to outputs spanning several orders of magnitude (e.g., atmospheric number densities, electron density) before z-scoring.
Neural Network Architecture:
- NRLMSISE, IGRF, and EGM: These models utilize compact Multi-Layer Perceptrons (MLPs) with HardSwish activation functions.
- IRI: Due to its complexity and the need to output constrained ion compositions (non-negative values summing to 1), a specialized "Multi-head MLP with normalization" architecture is used. This includes a shared trunk with Batch Normalization layers to stabilize deep training and separate output heads, with the ion composition head utilizing a Softmax activation to satisfy simplex constraints.
Training and Restoration: Models are trained using optimizers like AdaMax or AdamW with cyclical or one-cycle learning rate schedules. Post-inference, normalized predictions are restored to physical units via inverse transformations (e.g., exponentiation for log-transformed data).

Key Results
The study evaluates four surrogates against their respective single-threaded C++ reference implementations on an AMD Ryzen 5900X CPU and an NVIDIA RTX 4070 GPU.

Accuracy: All surrogates achieved sub-percent mean approximation errors within the tested input domains.
- NRLMSISE: Worst scalar Mean Absolute Percentage Error (MAPE) of 0.850% (Argon density).
- IRI: Worst scalar MAPE of 0.954% (Electron density); Mean Total Variation Distance (TVD) for ion composition was 0.24%.
- IGRF: Worst component MAPE of 0.732% (East component); Mean relative vector error of 0.021%.
- EGM: Worst component MAPE of 0.981% ( $\Delta a_x$ ); Mean relative vector error of 0.053%.
Performance (Speedup): Significant execution time reductions were observed, particularly for the IRI model, which is computationally expensive in its reference form.
- Single-threaded CPU vs. Reference C++: Speedups ranged from 4.3× (NRLMSISE) to 332× (IRI), 126× (IGRF), and 7.08× (EGM).
- GPU (RTX 4070) vs. Reference C++: Speedups were substantially higher, reaching 39.3× (NRLMSISE), 18,188× (IRI), 921× (IGRF), and 98.8× (EGM).
- The authors note that for very small batch sizes, fixed neural network overheads (model loading, framework initialization) can reduce the observed speedup, but these effects become negligible for large-scale batch processing.

Significance and Claims
The paper claims that neural-network surrogates can preserve high approximation accuracy (sub-percent mean error) while substantially reducing the computational cost of repeated geophysical model evaluations. The primary contribution is a validated workflow that transforms established, complex empirical models into efficient data-driven components.

The authors emphasize that this approach is particularly beneficial for applications requiring massive batch computations, such as satellite orbit propagation and observation modeling. While the surrogates are currently valid within the sampled training domains, the workflow provides a structured method for integrating these models into gradient-based computational workflows due to their differentiability. The study concludes that the magnitude of acceleration is primarily driven by the computational cost of the original reference implementation, with the most complex reference models (like IRI) yielding the highest speedups when replaced by compact neural networks.

The Cosmic Speedometer: Why We Need to Predict Space Weather Faster

The "Workaround" for Space Math

The Results: A Speed Demon with a Perfect Memory

Why This Matters

More like this