ConDiSim is a conditional diffusion model for simulation-based inference of complex systems with intractable likelihood. It approximates the posterior distribution using a denoising diffusion probabilistic model, consisting of a forward process that conditions observation data to add Gaussian noise to parameters and a backward process that removes noise. This method effectively captures the complex dependencies and multimodality within the posterior distribution. ConDiSim is evaluated on ten benchmark problems and two real-world problems, demonstrating effective posterior distribution approximation accuracy while maintaining computational efficiency and stability in model training.
