Machine Learning Reconstructs the Milky Way’s Gamma-Ray Sky from Planck Maps

A team of astrophysicists has shown that a machine learning model can predict the diffuse gamma-ray glow of the Milky Way using nothing but microwave and far-infrared maps of the sky, achieving an accuracy that rivals, and in the Galaxy’s crowded inner regions even surpasses, decades of painstaking physical modeling. The study, published in The European Physical Journal C, demonstrates that the nine frequency bands of the European Space Agency’s Planck satellite contain enough hidden information to reconstruct both the spatial pattern and the spectral shape of the high-energy emission that floods the Galactic sky.

The diffuse Galactic emission is the fog of high-energy astrophysics. It arises when cosmic rays, protons and electrons accelerated by sources such as supernova remnants, collide with interstellar gas, dust grains, magnetic fields, and the sea of starlight that permeates the Galaxy. Protons smashing into gas produce neutral pions that decay into gamma rays, while electrons scatter low-energy photons up to gamma-ray energies through inverse Compton processes and radiate synchrotron light in magnetic fields. Disentangling these components is notoriously difficult because their contributions overlap and depend on poorly constrained distributions of gas, radiation, and cosmic rays. Traditional models, such as the widely used GALPROP code, solve the cosmic-ray transport equation with detailed parameterizations, but they require extensive tuning and still struggle in the structurally complex inner Galaxy.

The research team, led by Xi Liu and colleagues at Sun Yat-sen University, took a deliberately different route. Rather than assuming a physical model from first principles, they trained two standard machine learning algorithms, Random Forest and K-Nearest Neighbors regression, to learn a direct mapping between Planck’s nine all-sky foreground maps, spanning 30 to 857 gigahertz, and the gamma-ray intensity predicted by the Fermi-LAT interstellar emission model across 28 energy bins from 50 megaelectronvolts to 814 gigaelectronvolts. Each Planck band traces a different physical regime: the lowest frequencies are dominated by synchrotron radiation from cosmic-ray electrons, intermediate bands mix free-free and spinning-dust emission, and the highest frequencies are dominated by thermal dust, a reliable tracer of interstellar gas.

The results are striking. In the 0.1 to 10 gigaelectronvolt range, where the diffuse gamma-ray sky is brightest, the models achieve coefficients of determination above 0.90, peaking at 0.96 at 687 megaelectronvolts. The predicted full-sky maps reproduce the major morphological features of the gamma-ray Galaxy, and the extracted spectral energy distribution of the inner Galactic plane matches the reference model closely. The team verified that the learned relationship is not a local accident: when a model trained on one Galactic hemisphere was tested on the opposite side, performance varied by no more than 6 percent regardless of how large a training region was used, indicating that the microwave-to-gamma-ray connection reflects a robust, large-scale property of the Milky Way rather than a statistical fluke of nearby structures.

Perhaps the most physically revealing result comes from asking which Planck frequencies matter most. When the high-frequency bands, dominated by thermal dust emission, were used alone, they achieved predictive power nearly identical to that of all nine bands in the 0.1 to 10 gigaelectronvolt range, while the low-frequency synchrotron channels performed markedly worse. This is exactly what one would expect if the diffuse gamma-ray emission at these energies is hadronically dominated: gamma-ray intensity is proportional to the product of gas column density and cosmic-ray density, and dust is an excellent proxy for the gas that serves as the target material. Above 10 gigaelectronvolts, however, the predictive power of the low-frequency bands rises, consistent with a growing leptonic contribution in which the synchrotron-emitting electron population also produces gamma rays through inverse Compton scattering. The analysis thus provides an independent, data-driven confirmation of the standard picture of Galactic gamma-ray production.

The residual maps, showing where the learned mapping breaks down, read like a tour of the Galaxy’s most enigmatic structures. Positive residuals appear toward the Magellanic Clouds and Centaurus A, where abundant gas and dust in nearby systems tempt the model to predict hadronic gamma-ray emission that the reference maps treat separately. Warm ionized gas complexes such as Barnard’s Loop in Orion and the Gum Nebula also show over-predictions, because free-free emission is subdominant across the Planck bands. On the other side, negative residuals trace Loop I and the North Polar Spur, hinting at inverse-Compton-dominated radiation fields or enhanced electron populations that submillimeter dust tracers cannot capture. Around the Galactic Center, the residuals mirror the shape of the Fermi Bubbles, the giant gamma-ray lobes whose faint, hard-spectrum microwave counterpart is too subtle for the model to disentangle without explicit priors.

The team also uncovered a curious hemispheric asymmetry: at high Galactic latitudes, the model systematically under-predicts the north while over-predicting the south by 10 to 30 percent on average. Because the discrepancy is spatially coherent rather than random, it may point to genuine physical differences between the two Galactic hemispheres, such as variations in cosmic-ray density or radiation fields. A segmented analysis of the inner Galaxy further showed that the asymmetry is concentrated within 30 degrees of the Galactic Center, likely linked to the Galactic bar and spiral-arm tangents, while the outer disk behaves with remarkable uniformity.

In a direct head-to-head comparison at roughly 4.3 gigaelectronvolts, the machine learning approach outperformed a GALPROP simulation tuned to the latest AMS-02 cosmic-ray data in the inner disk and Galactic Center region, achieving a coefficient of determination of 0.9465 and a mean absolute relative error of 14.7 percent, compared with 0.8624 and 22.3 percent for the physical model. GALPROP, by contrast, held a slight edge in the more homogeneous outer disk and in parts of the halo, confirming its reliability under simpler conditions. The authors emphasize that the two approaches are complementary: the data-driven model absorbs non-linear multi-frequency correlations that static gas distributions miss, while the physical simulation remains indispensable for interpretation and extrapolation.

An appendix to the study adds a compelling validation using raw Fermi-LAT photon counts rather than the smoothed emission model. As exposure accumulated over 40 weeks of mission data, the predictive performance climbed steadily and stabilized, and the negative residuals in the predicted maps turned out to coincide with 86 known gamma-ray point sources cataloged by Fermi, including pulsars, blazars, and a radio galaxy. Because the machine learning model learned the diffuse background purely from gas and dust maps without ever seeing point-source information, any compact gamma-ray emitter naturally surfaces as a localized deficit, offering an independent check on standard source-detection pipelines.

Beyond its immediate results, the work positions machine learning as a physically interpretable instrument rather than a black box. The learned mapping encodes real relationships between interstellar matter, radiation fields, and cosmic-ray processes, and its failures mark precisely the regions where conventional templates are incomplete or biased. The authors propose extending the framework with low-frequency radio surveys such as the Haslam 408-megahertz map, polarized microwave channels, and hydrogen-alpha data on ionized gas, and eventually integrating X-ray, ultra-high-energy gamma-ray, and neutrino observations. Such a multi-messenger, data-driven baseline could help separate standard emission from exotic components, sharpen constraints on cosmic-ray propagation, and, in an era when IceCube has detected Galactic neutrinos and LHAASO has measured PeV-scale diffuse emission, provide the empirical foundation that next-generation high-energy astrophysics will demand.

Subject of Research: Data-driven machine learning modeling of Galactic diffuse gamma-ray emission using multi-wavelength Planck observations

Article Title: Data-driven modeling of Galactic diffuse emission with multi-wavelength observations

Article References: Data-driven modeling of Galactic diffuse emission with multi-wavelength observations. (n.d.). https://doi.org/10.1140/epjc/s10052-026-16408-2

Image Credits: AI Generated

DOI: 10.1140/epjc/s10052-026-16408-2

Keywords: Galactic diffuse emission, machine learning, Planck satellite, Fermi-LAT, cosmic rays, gamma-ray astronomy, interstellar medium, inverse Compton, hadronic emission, GALPROP, Fermi Bubbles, multi-messenger astrophysics