Unsupervised Domain Adaptation for Constraining Star Formation Histories

TL;DR

Deriving the past star-formation history ($SFH$) of galaxies from present-day observations is hindered by domain shift between simulated and real spectra. The authors propose an unsupervised domain adaptation pipeline using KLIEP reweighting, with $SFH$ reduced via Kernel-PCA to $SFH_{kPCA}$ and a deep neural network trained on log-scaled SED features to predict $SFH_{norm}$ (and $SFH_{sum}$). Across three cross-domain experiments among SIMBA, Eagle, and IllustrisTNG, the approach improves predictions of both individual and global ($\Sigma$SFH) histories, as measured by multiple metrics (e.g., $RMSE$, $MAE$, $DTW$, $TDI$). While global trends are captured more reliably, the authors note persistent challenges in recovering highly stochastic SFH fluctuations and highlight smoothing as a practical step toward robust observational inferences, with future work on multi-source domain adaptation and physics-diverse simulations.

Abstract

The prevalent paradigm of machine learning today is to use past observations to predict future ones. What if, however, we are interested in knowing the past given the present? This situation is indeed one that astronomers must contend with often. To understand the formation of our universe, we must derive the time evolution of the visible mass content of galaxies. However, to observe a complete star life, one would need to wait for one billion years! To overcome this difficulty, astrophysicists leverage supercomputers and evolve simulated models of galaxies till the current age of the universe, thus establishing a mapping between observed radiation and star formation histories (SFHs). Such ground-truth SFHs are lacking for actual galaxy observations, where they are usually inferred -- with often poor confidence -- from spectral energy distributions (SEDs) using Bayesian fitting methods. In this investigation, we discuss the ability of unsupervised domain adaptation to derive accurate SFHs for galaxies with simulated data as a necessary first step in developing a technique that can ultimately be applied to observational data.

Figures (9)

• Figure 1: Radiation spectrums with corresponding star formation histories can be recorded from cosmological simulation to form a dataset $(X, y)$. This dataset is used to learn a machine learning model. Because of the domain shift between real radiation spectrums and simulated ones, domain adaptation is used to adapt the learned machine learning model to the real observations.
• Figure 2: Kernel density estimate plots of the $\log$ of flux densities (FD) for the first three features. The feature names at the top of each plot are the names of the filters, each centered at a different wavelength, in which photometry was simulated. We notice that for the three features shown here, all three simulations share the same support, which justifies our decision of using KLIEP.
• Figure 3: Global SFH predictions for the three experiments. The curves correspond to the sums over all SFH or predicted SFH.
• Figure 4: Comparison of SFH reduction approaches for SIMBA. On the left, the evolution of DILATE loss between the true sfh and the reconstructed signal after one of the three transformations: DWT, PCA, kernelPCA. On the right, reconstructed signal for one of the SIMBA SFH.
• Figure 5: KLIEP brings different domains closer by reweighing source samples. Top: Scatter plot of the first two PCA components in the input space. Bottom: KDE plots of log of the first two features.