The Explainable Universe working group at CosmicAI focuses on linking physical theories of cosmic structure formation with interpretable machine learning methods to enable robust, explainable inference from complex astronomical data. This effort is organized around two complementary thrusts: (1) the generation of large, systematically varied suites of cosmological simulations to probe the nature of Dark Matter and cosmic structure formation, and (2) the development of explainable AI methods for scientific inference under realistic, imperfect data conditions.

On the astrophysics side, the group is using commonly employed codes, including AREPO, RAMSES, and GIZMO, to generate controlled cosmological simulation suites. In distinct contrast to previous studies (e.g., like Illustris, IllustrisTNG, Simba, FIREBox, etc.), these simulation suites include thousands of simulations that systematically vary the input physical assumptions (e.g., the properties of Dark Matter) to create maps linking changed physical assumptions to identifiable impacts on galaxy formation and galaxy properties. By producing a broad library of realizations, we can trace how fundamental physics shapes observable structures in the Universe. This work is closely intertwined with the broader DREAMS collaboration and will scale toward larger, more diverse parameter spaces in upcoming runs.

On the AI side, the group develops simulation-based inference (SBI) pipelines that pair these simulation suites with machine learning models that draw probabilistic, interpretable conclusions about underlying cosmological parameters. Methods under development include graph neural networks for learning structural relationships between galaxies and their environments, uncertainty-aware inference models to quantify confidence, and explainability techniques such as saliency mapping and feature attribution to identify which physical signatures most strongly constrain model parameters. The emphasis is on explainability by design — ensuring that models provide insight into why they reach specific conclusions.

A central challenge in cosmology is that observational data are incomplete and noisy, and are subject to biases that can obscure underlying physics. Our inference framework explicitly addresses this through robustness testing, domain adaptation, and uncertainty calibration. These steps allow us to draw trustworthy scientific inferences even when real data deviate from idealized training sets. This capability is particularly crucial for leveraging upcoming large-scale survey missions, which will provide unprecedented but complex data.

Looking ahead, the Explainable Universe working group will expand both sides of this pipeline: scaling astrophysical simulation suites to encompass more physical models and initial conditions, and advancing AI frameworks to support explainable, uncertainty-aware inference at survey scale. This approach aims to turn astronomical observations into transparent, scientifically grounded constraints on the nature of Dark Matter and cosmic structure formation.