A Demonstration of a Neural Network as a Bridge Between Galaxy Simulations and Surveys

This paper demonstrates that the stellar masses of galaxies in the Galaxy and Mass Assembly (GAMA) survey, originally derived via stellar population synthesis modelling, can be accurately predicted using only their absolute magnitudes and colour indices. A central contribution of this work is the demonstration that this long-standing inference problem can be solved using an exceptionally simple machine-learning model: a fully connected, feed-forward artificial neural network with a single hidden layer. The network is trained exclusively on synthetic galaxies generated by the SHARK semi-analytic model and is shown to transfer effectively to real observations. Across nearly 3.5 dex in stellar mass, the predicted values closely track the GAMA SED-derived masses, with a typical scatter of ~0.131 dex. These results demonstrate that complex deep-learning architectures are not a prerequisite for robust stellar mass estimation, and that simulation-trained, lightweight machine-learning models can capture the dominant physical information encoded in broad-band photometry. The method is further applied to 17,006 GAMA galaxies lacking SED-derived masses, with photometric uncertainties propagated through the network to provide corresponding error estimates on the inferred stellar masses. Overall, this work establishes a computationally efficient and conceptually transparent pathway for simulation-to-observation transfer learning in galaxy evolution studies.

💡 Research Summary

Estimating galaxy stellar masses is a cornerstone of extragalactic astronomy, yet the quantity is not directly observable and must be inferred from spectral energy distribution (SED) fitting. Traditional SED‑based methods rely on a suite of assumptions about star‑formation histories, metallicities, dust attenuation, and the initial mass function, leading to systematic uncertainties of order 0.2–0.3 dex. In parallel, cosmological galaxy formation simulations such as the SHARK semi‑analytic model predict stellar mass as a fundamental output and naturally encode relationships between broadband photometry and stellar mass. The central question addressed in this work is whether a model trained solely on simulated galaxies can be transferred to real survey data to recover accurate stellar masses without any exposure to observations during training.

The authors construct a very simple machine‑learning architecture: a fully connected feed‑forward artificial neural network (ANN) with a single hidden layer. Input features consist of 24 absolute magnitudes and colour indices spanning the UV to mid‑infrared (GALEX FUV/NUV, SDSS u,g,r,i,z, VISTA Y,J,H,Ks, WISE W1,W2) that are available for both the SHARK mock catalogue and the GAMA DR4 survey. The network is trained on millions of SHARK galaxies using mean‑squared error loss, Adam optimisation, L2 regularisation, and early stopping to avoid over‑fitting. After retraining with the reduced feature set (the unavailable UV colour FUV‑NUV and MIR colour W1‑W2 are omitted), the ANN achieves an RMS error of ≈0.117 dex on a held‑out SHARK validation set.

For the observational test, the authors assemble a clean GAMA sample of 71 171 galaxies with reliable spectroscopic redshifts (NQ > 2) and high‑quality photometry. They ensure that every input lies within the range spanned by the SHARK training data; objects falling outside the SHARK limits in FUV‑NUV or W1‑W2 are excluded, and a small number of galaxies with marginally out‑of‑range magnitudes are removed. The ANN is then applied directly to the GAMA photometry to predict stellar masses.

When compared to the GAMA SED‑derived masses (based on Bruzual & Charlot 2003 models and a Chabrier IMF), the ANN predictions follow the one‑to‑one line across a 3.5 dex mass range (10⁸․⁵–10¹² M⊙). A second‑order polynomial fit reveals a modest systematic offset of up to ~0.1 dex, which can be removed with a simple bias correction. After correction, the residual distribution has a half‑width of the 16–84 percentile range of 0.135 dex, indicating a scatter comparable to, or slightly better than, typical SED‑based uncertainties. The intrinsic scatter of the ANN itself (≈0.13 dex) dominates the error budget, confirming that the model generalises well despite never having seen real data during training.

To demonstrate practical utility, the trained ANN is applied to 17 006 GAMA galaxies that lack SED‑derived masses. Photometric uncertainties are propagated by perturbing each flux by ±½ its reported error, evaluating the network three times per galaxy, and extracting central values with upper and lower bounds. The resulting stellar masses show a tight linear relation with WISE W1 absolute magnitude, as expected because W1 traces the old stellar population. Propagated photometric uncertainties are typically ±0.05 dex, much smaller than the ANN’s intrinsic scatter; a conservative total uncertainty of ≈0.18 dex is therefore adopted.

The paper discusses several limitations: (i) the exclusion of the FUV‑NUV colour removes a strong UV attenuation diagnostic, slightly degrading performance; (ii) SHARK, like any semi‑analytic model, may not capture the full diversity of real galaxy physics (e.g., extreme AGN feedback, environmental quenching); (iii) the method has not been tested on galaxies with colours far outside the training domain. Nonetheless, the study convincingly shows that a minimalist neural network can capture the dominant photometric‑mass information encoded in simulations and transfer it to observations with high fidelity.

In summary, the authors provide a proof‑of‑concept that simulation‑trained, lightweight ANNs can deliver accurate stellar mass estimates for large survey samples using only broadband magnitudes and colours. The approach yields a typical scatter of ≤0.13 dex across a broad mass range, comparable to traditional SED fitting but at a fraction of the computational cost. This work opens the door to broader applications of transfer learning in galaxy evolution, suggesting that more complex deep‑learning architectures are unnecessary for many fundamental inference tasks.