Adaptation to a new domain. Left: Train ensemble of five clustering NNs with BAHAMAS, DARKSKIES-0, and DARKSKIES-0.2 as known (solid blue and red shades) and DARKSKIES-0.1 as unknown (hatched black). We show the combined PDF of log(σ_DM/m) from the five runs. We find a cross-section of ∼0.14 cm² g⁻¹ for DARKSKIES-0.1, within 1σ of 0.1 cm² g⁻¹. Right: Overlap confidence (left axis, solid blue), estimated σ_DM/m (right axis, dashed purple), and target σ_DM/m (right axis, dotted black) for DARKSKIES-0.1 as we progressively add known simulations during training. We start with BAHAMAS-0 and BAHAMAS-SIDM, then add the BAHAMAS-AGN simulations, DARKSKIES-0, and finally DARKSKIES-0.2. We see that the overlap confidence closely follows the convergence of DARKSKIES-0.1 onto the target σ_DM/m with an overlap of 70% resulting in DARKSKIES-0.1 being within 1σ of the target σ_DM/m.