Conditional Diffusion Surrogates for Modified-Gravity Cosmological Simulation

Scientific inference in modified-gravity cosmology requires large ensembles of nonlinear matter-density fields, but producing these fields with numerical simulations is computationally expensive. We develop a class-conditional denoising diffusion probabilistic model as a generative surrogate for high-fidelity modified-gravity cosmological simulations. We train a single model conditioned on four discrete values of the modified-gravity parameter $f_{R_0}$. We evaluate the generated fields using power spectra, class ratios, and peak counts, which probe clustering, gravity-dependent trends, and non-Gaussian structure. Across all four gravity regimes, the model reproduces the simulated power spectra and preserves the scale-dependent differences between $f(R)$ classes. The generated fields also recover the peak-count distribution over the well-sampled range. These results provide a proof of concept for diffusion-based emulation of modified-gravity density fields. They support future extensions to three-dimensional fields and continuous parameter conditioning, with the longer-term goal of accelerating mock generation and simulation-based inference.