The era of large-scale astronomical surveys demands innovative approaches for the rapid and accurate analysis of vast spectral datasets. A promising direction to address this challenge lies in the synergy between physics-based simulations and machine learning. Astronomy, like many other scientific disciplines, relies heavily on simulations to encode our physical understanding of complex processes, generate hypotheses, design experiments, and explore causal relationships. Increasingly, this knowledge is embedded in forward models that produce realistic synthetic spectra, images, and data-cubes. A central and long-standing challenge, however, is the solution of the “inverse problem”: inferring the underlying physical parameters from observational data. In this talk, I will present a powerful framework to tackle this problem based on simulation-based inference (SBI) accelerated by invertible neural networks. Thanks to their amortised nature, these methods offer significant advantages over traditional Bayesian approaches: once trained, they enable fast inference and naturally yield complex, multi-modal posterior distributions conditioned on observables. I will show preliminary results based on simulated spectra for upcoming instruments, along with ongoing SBI applications aimed at probing the interstellar medium in galaxies. Finally, I will discuss key open challenges, including model interpretability and domain adaptation, with particular emphasis on bridging the gap between synthetic training data and real observations.
Decoding galaxy properties with machine learning and simulation based inference