SBI⁺⁺: Flexible, Ultra-fast Likelihood-free Inference Customized for Astronomical Applications

Abstract

Flagship near-future surveys targeting 108–109 galaxies across cosmic time will soon reveal the processes of galaxy assembly in unprecedented resolution. This creates an immediate computational challenge on effective analyses of the full data set. With simulation-based inference (SBI), it is possible to attain complex posterior distributions with the accuracy of traditional methods but with a >104 increase in speed. However, it comes with a major limitation. Standard SBI requires the simulated data to have characteristics identical to those of the observed data, which is often violated in astronomical surveys due to inhomogeneous coverage and/or fluctuating sky and telescope conditions. In this work, we present a complete SBI-based methodology, SBI ++ , for treating out-of-distribution measurement errors and missing data. We show that out-of-distribution errors can be approximated by using standard SBI evaluations and that missing data can be marginalized over using SBI evaluations over nearby data realizations in the training set. In addition to the validation set, we apply SBI ++ to galaxies identified in extragalactic images acquired by the James Webb Space Telescope, and show that SBI ++ can infer photometric redshifts at least as accurately as traditional sampling methods—and crucially, better than the original SBI algorithm using training data with a wide range of observational errors. SBI ++ retains the fast inference speed of ∼1 s for objects in the observational training set distribution, and additionally permits parameter inference outside of the trained noise and data at ∼1 minute per object. This expanded regime has broad implications for future applications to astronomical surveys. (Code and a Jupyter tutorial are made publicly available at https://github.com/wangbingjie/sbi_pp.)

The UNCOVER Survey: A First-look HST+JWST Catalog of Galaxy Redshifts and Stellar Population Properties Spanning 0.2 ≲ z ≲ 15

“…Thus, the difference between the emulator and FSPS results are in general well described by the reported uncertainties in the catalog. An additional valuable, independent check on the fidelity of the inferred parameters can likely be obtained via an alternative inference technique, since the latter is affected by different systematics (Wang et al 2023d), or by comparing to additional spectra. This approach will be examined in the next generations of stellar populations catalogs.…”

“…The sampling method used is nested sampling (Skilling 2004), which is better suited to sample multimodal posteriors than other traditional techniques such as Markov Chain Monte Carlo (Goodman & Weare 2010). However, it has also been shown that nested sampling does not always accurately sample the global minimum when fitting for galaxy redshifts (Wang et al 2023d). Our finding is consistent with…”

Stochastic Modeling of Star Formation Histories. III. Constraints from Physically Motivated Gaussian Processes

“…In practice, observational noise due to instrumental effects and measurement uncertainties might necessitate larger samples. While a detailed modeling of the noise considering wavelength coverage, resolution, extent of correlated noise across wave bands, and overall S/N is outside the scope of this work, including an accurate noise model in the forward modeling "simulator" in the SBI while training can significantly improve its robustness, especially with newer methods (e.g., SBI++; Wang et al 2023) being developed.…”

PopSED: Population-level Inference for Galaxy Properties from Broadband Photometry with Neural Density Estimation

“…Using traditional SED fitting methods, analyzing 10 5 galaxies will take up to 2 × 10 6 CPU hr. Even with the development of accelerated SED fitting (e.g., Alsing et al 2020;Hearin et al 2023;Khullar et al 2022;Wang et al 2023), an analysis of 10 5 galaxies will still take up to ∼10 3 GPU hr. POPSED is able to recover the posterior of the population distribution for ∼10 5 galaxies within ∼10 GPU hr, 100 times faster than the SBIbased methods.…”

“…This problem has been partially mitigated by recent developments in accelerating SED modeling by emulating the SPS models with neural networks (Alsing et al 2020), building differentiable SPS models with a high-performance library (Hearin et al 2023), and speeding up the sampling by using amortized simulation-based inference (SBI; Khullar et al 2022;Wang et al 2023). However, some of these methods need sophisticated training and are costly to retrain when adapting to different SPS models or noise properties.…”