⚛️ phenomenology

Fast Surrogate for the Earth Matter Effect on Solar Neutrinos

This paper introduces a deep-learning surrogate model based on a residual convolutional network that rapidly approximates the Earth matter effect on solar neutrinos with 2% accuracy and a 60-fold speed-up compared to traditional numerical calculations, offering a flexible tool for both standard and new-physics analyses.

Original authors: Saeed Ansarifard

Published 2026-09-23

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Saeed Ansarifard

Original paper licensed under CC BY 4.0 (http://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

Sunlight is not just light; it is a flood of tiny, ghostly particles called neutrinos, born in the nuclear furnace at the heart of our star. These particles stream outward in vast numbers, passing through the Earth and our bodies without us ever feeling them. To understand the universe, physicists study how these particles change their identity as they travel. A solar neutrino might start its journey as an electron-neutrino, but by the time it reaches a detector on Earth, it may have shifted into a different type. This transformation, known as oscillation, depends on the particle's energy and the material it travels through. When these neutrinos pass through the Earth itself, the density of the planet's interior subtly alters their behavior, a phenomenon known as the Earth matter effect. Calculating exactly how this happens is crucial for interpreting data from modern experiments, but doing so requires solving complex equations for every possible path a neutrino might take and for every possible value of the physical constants that govern their behavior. This process is so computationally heavy that it can slow down the entire analysis of experimental data, turning what should be a quick calculation into a days-long wait.

A researcher at the Institute for Research in Fundamental Sciences in Iran has developed a new tool to bypass this bottleneck. Instead of solving the heavy equations every time a calculation is needed, the team trained a computer program to learn the pattern of the Earth's influence on these particles. They created a digital model that acts as a fast substitute, or surrogate, for the slow, traditional math. The model was built using a type of artificial intelligence known as a deep-learning network. The researchers fed the network thousands of examples of how neutrinos behave, generated by running the full, slow calculations on a grid of different physical parameters. Specifically, they focused on two key numbers that describe how neutrinos mix and change: the angle of mixing between two types of neutrinos and the difference in the squares of their masses. The network learned to look at these two numbers and instantly predict a detailed map showing how the neutrinos would change as they traveled through the Earth at different energies and from different angles.

The result is a system that can predict the outcome of these complex interactions in a fraction of the time it takes the old method. In tests run on a standard laptop computer, the new model produced its predictions about sixty times faster than the direct numerical calculation. Despite this incredible speed, the model remained highly accurate, with a point-by-point error of only about two percent compared to the slow, precise method. The researchers verified that the model could reproduce the intricate patterns of neutrino behavior across a wide range of energies and travel paths, from those skimming the Earth's surface to those cutting straight through its core. Because the model was designed to learn only the specific physics of how neutrinos move through the Earth, it remains flexible. If scientists discover new physics that changes how neutrinos are created before they reach the Earth, this model can still be used without needing to be retrained, as long as the journey through the Earth itself remains unchanged.

The team tested their creation by comparing its predictions against the results of the traditional, slow calculations for a set of parameters it had not seen during training. The surrogate model successfully recreated the complex shapes and variations of the neutrino behavior maps, capturing the subtle ways the Earth's density affects the particles. The accuracy was particularly high for neutrinos traveling at certain angles, and the average error across all conditions stayed well within the limits required for current scientific analyses. This work does not claim to have solved the mystery of neutrinos, but it provides a powerful new way to handle the massive amounts of data coming from upcoming experiments. By replacing a slow, repetitive calculation with a fast, learned prediction, the researchers have removed a significant barrier to studying the properties of these elusive particles. The code and the data used to build this model are now available to other scientists, allowing the broader community to use this fast approximation in their own searches for the secrets of the solar neutrinos.

Technical Summary: Fast Surrogate for the Earth Matter Effect on Solar Neutrinos

Problem Statement
The propagation of solar neutrinos through the Earth requires solving the neutrino evolution equation in a medium with a density profile that varies along the trajectory. While analytic approximations exist for low-density regimes or averaged observables, they are insufficient for modern analyses requiring fine energy and time binning. Direct numerical solutions, though accurate, are computationally expensive when repeated over broad parameter spaces and large numbers of trajectories, as required for global fits, likelihood evaluations, and Monte Carlo analyses. The computational burden grows with upcoming experiments demanding higher resolution, necessitating a fast and accurate surrogate model to approximate the Earth matter effect without solving the Schrödinger equation for every evaluation.

Methodology
The author proposes a deep-learning surrogate model implemented as a parameter-conditioned convolutional neural network with a ResNet-based encoder–decoder architecture. The model is designed to learn the mapping between physical oscillation parameters and the resulting Earth-matter transition probabilities in a two-dimensional energy–trajectory plane.

Physics Reduction: The model focuses on the electron-neutrino survival probability ( $P_{ee}$ ). The author demonstrates that $P_{ee}$ is effectively independent of the CP-violating phase $\delta_{CP}$ and the mixing angle $\theta_{23}$ . Furthermore, $\theta_{13}$ is fixed to its global best-fit value due to its high precision and negligible impact on $P_{ee}$ within its allowed range. The mass-squared difference $\Delta m^2_{31}$ is also fixed, as its associated oscillation length is orders of magnitude shorter than the Earth's radius, causing the phase to average out. Consequently, the relevant evolution operator depends only on two free parameters: the mixing angle $\theta_{12}$ and the mass-squared difference $\Delta m^2_{21}$ .
Network Architecture:
- Input: The network takes a vector of the two oscillation parameters, $x = (\theta_{12}, \Delta m^2_{21})$ .
- Target: The output is a 2D image (192 $\times$ 192 grid) representing the three transition probability components ( $I_1, I_2, I_3$ ) as functions of neutrino energy ( $E_\nu \in [5, 20]$ MeV) and zenith angle ( $\eta \in [0, 95^\circ]$ ).
- Structure: The model employs a dual-encoder approach during training. An image encoder (ResNet-based) maps the numerically computed target images into a compact latent space. A parameter encoder (MLP) maps the physical input parameters into a latent representation of identical dimensionality. A residual decoder then maps the parameter-derived latent representation back to the predicted image.
- Training Objective: The network is trained using a combined loss function: a direct image-space reconstruction loss ( $L_{image}$ ) and a latent-space matching loss ( $L_{latent}$ ) that forces the parameter-derived latent vector to match the image-derived latent vector. This ensures the parameter encoder learns a representation consistent with the manifold of numerical solutions.
Dataset: The training set consists of 441 parameter pairs sampled from a $21 \times 21$ grid covering $\theta_{12} \in [30^\circ, 40^\circ]$ and $\Delta m^2_{21} \in [6, 9] \times 10^{-5}$ eV $^2$ . The target images are generated using the PEANUTS package. The dataset is split into 80% training and 20% validation.

Key Results

Accuracy: The surrogate model achieves a pointwise relative accuracy of approximately 2% for the reference electron-neutrino survival probability. The root-mean-square error (RMSE) averaged over the propagation angle remains below $\sim 0.01$ across the energy range, corresponding to an average prediction error of approximately 3%. The accuracy is highest for trajectories near the horizon ( $\eta \approx 90^\circ$ ) and slightly lower for core-crossing trajectories ( $\eta \approx 0^\circ$ ), though these localized errors are expected to average out in time-integrated observables.
Performance: In a test using an Apple M1 laptop CPU, the surrogate provides a speed-up of approximately a factor of 60 compared to direct numerical calculation. Generating around 100 predicted maps takes approximately 2 seconds, whereas the equivalent numerical calculation takes over 120 seconds. The author notes that speed-ups on GPU hardware are expected to be even larger.
Generalization: The model successfully reproduces the characteristic energy and angular dependence of the survival probability across different regions of the parameter space, including the central values and the boundaries of the training grid.

Significance and Claims
The paper claims that this surrogate model offers a flexible and computationally efficient approach for solar-neutrino analyses, particularly for tasks requiring evaluations at a vast number of parameter points, such as global fits and likelihood scans.

A distinct feature of the framework is its modularity. Since the network learns only the Earth-crossing propagation (the Hamiltonian inside the Earth), it can be applied without retraining to new-physics scenarios that modify the neutrino state before it reaches the Earth (e.g., long-range interactions affecting mass-eigenstate fractions at the surface) but leave the propagation Hamiltonian unchanged. Conversely, for scenarios involving non-standard neutrino interactions (NSI) that alter the propagation Hamiltonian, the surrogate can be retrained on new numerical samples generated with the modified potential, with the new physics parameters included as additional inputs to the parameter encoder.

The author emphasizes that the current accuracy reflects interpolation within a densely sampled grid. While the trained range covers the region allowed by current global fits, explicit testing of extrapolation beyond these bounds is identified as a necessary next step for broader applicability. The code and simulation data are made publicly available to facilitate further development and application.

Technical Summary: Fast Surrogate for the Earth Matter Effect on Solar Neutrinos

More like this