DiffDA is a novel machine learning (ML) data assimilation method designed for weather forecasting and climate modeling. It leverages a denoising diffusion model to assimilate atmospheric variables using predicted states and sparse observations. The method uses the GraphCast neural network as its backbone, which is adapted to handle the input and output shapes required by the diffusion model. Through experiments on simulated observations from the ERA5 reanalysis dataset, DiffDA achieves high-resolution assimilation (0.25° resolution) and demonstrates that assimilated data can be used for forecasts with a lead time loss of up to 24 hours. The method also enables the creation of reanalysis datasets using autoregressive data assimilation. Key contributions include the development of a high-resolution ML data assimilation method, the creation of data assimilation cycles with an ML weather forecasting model, and the flexibility of using a pretrained ML forecast model as the backbone of the diffusion model. The experimental results validate the effectiveness of DiffDA, showing that assimilated data converge to observations as the number of observed data points increases and that the method can produce accurate forecasts with a limited number of sparse observations.DiffDA is a novel machine learning (ML) data assimilation method designed for weather forecasting and climate modeling. It leverages a denoising diffusion model to assimilate atmospheric variables using predicted states and sparse observations. The method uses the GraphCast neural network as its backbone, which is adapted to handle the input and output shapes required by the diffusion model. Through experiments on simulated observations from the ERA5 reanalysis dataset, DiffDA achieves high-resolution assimilation (0.25° resolution) and demonstrates that assimilated data can be used for forecasts with a lead time loss of up to 24 hours. The method also enables the creation of reanalysis datasets using autoregressive data assimilation. Key contributions include the development of a high-resolution ML data assimilation method, the creation of data assimilation cycles with an ML weather forecasting model, and the flexibility of using a pretrained ML forecast model as the backbone of the diffusion model. The experimental results validate the effectiveness of DiffDA, showing that assimilated data converge to observations as the number of observed data points increases and that the method can produce accurate forecasts with a limited number of sparse observations.