Abstract: Many data-driven surrogate models have been proposed to alleviate the computational burden of computational fluid dynamics (CFD) simulations. However, an adequate amount of training data is required for building data-driven surrogate models. Due to the huge consumption of simulation calculations, it is difficult to obtain the data required for training surrogate model. To remove the obstacle, this paper proposes an Interpolation Cold Diffusion Model (ICDM)-based data-driven surrogate modeling approach that only uses limited amount of CFD simulation data. The proposed method transforms the prediction of the species mass fraction distributions using boundary conditions into an image-to-image translation task, which translates the source domain species mass fraction distributions to the target domain species mass fraction distributions according to the target domain boundary conditions. Different from the original Denoising Diffusion Probabilistic Model (DDPM) that adds random Gaussian noise, ICDM interpolates between distributions in diffusion process, which can provide more information for reducing the needed amount of the training data. The proposed approach is demonstrated via 2D methane combustion experiment simulated by the CFD. By the training in 40 data points, this model produces overall normalized mean absolute errors of 1.89% for CH₄ mass fraction, which is even better than some models constructed by other generation models with 800 data points, and 7 times faster than CFD model. The proposed approach can utilize only a small amount of training data, quickly building an accurate and computationally cheaper CFD surrogate model, providing a basis for deploying virtual sensors, visualizing internal reaction processes, conducting rapid reaction analysis, and exploring the optimization of reaction processes.