ASTROLABE Delivers Grid-Free Particle Emulation of Lambda-CDM Cosmology via Physical Flow Matching

The model replaces standard noise priors with COLA-based initial coordinates and enforces momentum and mass conservation along the flow trajectory. A field-particle hybrid transformer lets every particle cross-attend to multiscale field tokens, keeping memory linear in particle count. This architecture directly outputs particle positions and velocities, allowing any downstream observable to be computed without interpolation artifacts that plague grid-based emulators.

Tests on Lambda-CDM initial conditions reproduced the nonlinear matter power spectrum to high accuracy across scales, halo mass functions within a few percent of reference N-body runs, and realistic velocity and density profiles inside halos. Out-of-distribution performance on dark-acoustic-oscillation initial conditions indicates the physical prior helps the model generalize beyond training cosmologies. Fine-tuning experiments successfully scaled the same weights to emulate a full TNG50-3-dark volume.

Compared with earlier ML emulators limited to summary statistics or low-resolution grids, ASTROLABE removes the resolution ceiling that has constrained systematic dark-matter model scans. The approach connects to recent advances in continuous normalizing flows and physically constrained generative models, suggesting a path toward baryonic inclusion once training data volume grows. Remaining gaps include absence of hydrodynamics and limited validation at the highest resolutions demanded by upcoming surveys.

Next steps center on incorporating baryonic physics modules and expanding training to varied dark-sector models, which would enable rapid parameter scans previously blocked by simulation cost.