Decoding stellar barcodes with Lux

The large volume of spectroscopic data available now and from near-future surveys will enable high-dimensional measurements of stellar parameters and properties. Current methods for determining stellar labels from spectra use physics-driven models, which are computationally expensive and have limitations in their accuracy due to simplifications. While machine learning methods provide efficient paths toward emulating physics-based pipelines, they often do not properly account for uncertainties and have complex model structure, both of which can lead to biases and inaccurate label inference. Here we present Lux: a data-driven framework for modeling stellar spectra and labels that addresses prior limitations. Lux is a generative, multioutput, latent-variable model framework built on JAX for computational efficiency and flexibility. As a generative model, Lux properly accounts for uncertainties and missing data in the input stellar labels and spectral data and can either be used in probabilistic or discriminative settings. Here, we present several examples of how Lux can successfully emulate methods for precise stellar label determinations for stars ranging in stellar type and signal-to-noise ratio from the APOGEE survey. We also show how a simple Lux model is successful at performing label transfer between the APOGEE and GALAH surveys. Lux is a powerful new framework for the analysis of large-scale spectroscopic survey data. Its ability to handle uncertainties while maintaining high precision makes it particularly valuable for stellar survey label inference and cross-survey analysis, and the flexible model structure allows for easy extension to other data types.