In this talk, three aspects of fusion device design will be discussed in which a hierarchy of models are available: turbulent transport, electromagnets, and confinement of energetic particles. In the area of turbulent transport, surrogates are derived from a dataset of more than 200,000 nonlinear gyrokinetic simulations of ion-temperature-gradient turbulence in diverse non-axisymmetric geometries. Analysis with interpretable machine-learning methods identifies flux-surface compression in regions of bad curvature, followed by geodesic curvature, as the most important geometric predictors. These results align with some prior analytical work but also pose questions for theory. For electromagnet design, a reduced model is presented for the internal magnetic field, self-force, and self-inductance of coils in which singular filament integrals are regularized, so that rapid 1D calculations accurately reproduce high-fidelity results. The differentiable reduced model is demonstrated in gradient-based optimization. For energetic particle confinement, data-driven parameterizations of stellarator shapes are presented that provide natural bounds, improved scaling, and dimensionality reduction for Bayesian optimization with high-fidelity calculations in the optimization loop. These optimization yield configurations with excellent fast-particle confinement despite having significant deviations from quasisymmetric, omnigenous, and quasi-isodynamic field patterns. Together, these examples present several areas where reduced, data-driven, and high-fidelity physics models could be coupled in fusion device design.