To What Extent Are Star Cluster Ages Encoded in Their Environments? Exploring the Spatial Distribution of Age-Related Information with PHANGS-HST Imaging and Convolutional Neural Networks

TL;DR

It is demonstrated that age-predictive environmental cues are encoded at a level detectable by machine-learning and recoverable from broadband imaging, which establishes a path for using new techniques to connect image-based age inference to the physical evolution of cluster environments.

Abstract

The environments around star clusters evolve as stellar feedback reshapes the interstellar medium and dynamical processes reorganize the structure of the surrounding stellar field. As approximately single-age populations, star clusters can serve as clocks to trace these environmental changes. In this exploratory study, we test whether convolutional neural networks (CNNs) can identify age-dependent changes in cluster environments. We take cluster ages as given from basic SED fitting of five-band UV-optical aperture photometry from the PHANGS (Physics at High Angular resolution in Nearby GalaxieS) HST survey. We first show that CNNs can be trained on image cutouts centered on clusters to recover ages directly from imaging. This demonstration provides the foundation for this study, which examines whether the information used by CNNs to predict age is coherent and physically meaningful. We perform controlled image occlusion experiments as an explainable AI method. These show that the CNNs extract age-predictive environmental cues in the absence of cluster light and when information on SED shape is removed by combining the five filters into one image. We find that reliance on environmental information increases at the youngest (<10 Myr) and oldest (>1 Gyr) ages, where clusters can exhibit similarly red colors. Our results are consistent with the long-recognized picture that cluster environments evolve systematically with age. We demonstrate that this information is encoded at a level detectable by machine-learning and recoverable from broadband imaging. This establishes a path for using new techniques to connect image-based age inference to the physical evolution of cluster environments.

Figures (8)

• Figure 1: Examples of the 112$\times$112 pixel HST image cutouts used in the CNN analysis. Three star clusters are shown, each centered in the cutouts, with the full set of five band NUV U B V I images. The examples are chosen to span the distance range of the PHANGS galaxy sample. The red circle marks the adopted photometric aperture (radius of 4 pixels or 0$.\!\!^{\prime\prime}$14). After applying a correction for light lost outside the aperture deger22, this photometry is used to derive the cluster ages for the PHANGS cluster catalog through spectral energy distribution fitting as described in turner21. These cluster ages are the ones used to train the CNNs in this work.
• Figure 2: CNN age-prediction performance for two training samples: (left) training and evaluation on the full PHANGS-HST cluster sample and (right) training and evaluation with the youngest and oldest clusters removed. The top panels show predicted versus reference ages, demonstrating that the CNN can indeed recover the SED-fit ages from the imaging but exhibits systematic residuals at the extreme ages. The middle panels plot residuals as a function of true age, highlighting the overestimation of very young clusters and the underestimation of very old clusters in the full-sample model. The bottom panels display training and validation mean-squared error (MSE) as a function of the number of passes through the training set. Together, these comparisons confirm the feasibility of age recovery from imaging.
• Figure 3: Blackout experiments to assess the spatial localization of age-predictive information in image cutouts centered on star clusters. A circular mask of increasing diameter is applied concentrically over each image, progressively removing central and surrounding regions. The $x$-axis shows the blackout diameter (in pixels), while the $y$-axis reports the mean scatter (dex) of five independently trained models for training (Tr.), validation (Vl.), and test (Ts.) datasets. The top row corresponds to the 5-filter HST input, and the bottom row to the stacked single white-light image. Left panels use the full cluster sample, while right panels exclude extreme-age systems. Dashed horizontal lines at the top of each panel indicate the fixed-guess baseline (non-learned average prediction). The black silhouettes illustrate representative blackout masks at different diameters. The different color regions mark the transition from cluster-dominated zones (green) to environment-dominated zones (orange). The black curve at the bottom of each panel quantifies the relative predictive power retained in the visible pixels, computed $(\mathrm{baseline}_{\mathrm{test}}^{2} - \mathrm{curve}_{\mathrm{test}}^{2}) / \mathrm{baseline}_{\mathrm{test}}^{2}$. These experiments quantitatively demonstrate that age-predictive information is present in the environment surrounding clusters.
• Figure 4: Performance comparison of 8 cases for cluster age prediction, using the full range of cluster ages (log(age) from 6.0 to 10.0). Case 1 includes unmasked cutouts in all 5 HST filters and provides the maximum information on cluster color, morphology, and environment, as reported in Section \ref{['sec:feasibility']} and Figure \ref{['fig:true_vs_pred']}. Case 2 uses a stacked white-light image, retaining only morphology and environment. Cases 3 and 4 use the stacked image with the environment removed using 24-pixel and 12-pixel diameter outer masks, respectively, thereby focusing the model on the cluster light. Cases 5 and 6 apply the mask inverse, masking out the cluster center with 12-pixel and 24-pixel diameters, respectively, thereby focusing the model on the environment, as reported in Section \ref{['sec:innermask']} and Figure \ref{['fig:blackout_experiments']}. Case 7 uses the same outside-24-pixel environment input as Case 6, but with randomized reference ages during training and validation. This prevents the network from learning any physically meaningful correlations, causing it to converge to predictions near the mean age of the training set and to exhibit performance comparable to the null baseline. Case 8 shows the performance of the null predictor, providing a lower bound for comparison. The bars show the mean scatter (dex) of log(age), which is adopted as the performance metric. Each case is based on 5 independently trained CNN models, with error bars showing variability across models. All results refer to the test set.
• Figure 5: Similar to Figure \ref{['fig:five_cases']}, but now also showing results for models trained on a restricted set of reference ages (log(age) = 7.0–9.5), in which the youngest and oldest clusters have been excluded (right panel of bar charts). Cases 1 and 4–8 are reproduced from Figure \ref{['fig:five_cases']} in the left panel of bar charts. Two additional occlusion experiments using the full five-filter input are added (Cases 2 and 3 in this figure), applying outer and inner masking, respectively.