Understanding and predicting the Earth system, especially extreme events, have profound impacts on people, societies and ecosystems. Traditional physics-based models like E3SM are computationally demanding and struggle with large ensemble simulations needed for quantifying prediction uncertainties and capturing low-probability events. They also face challenges in integrating multitype and multiscale data due to limited processes and mesh-grid configurations. While machine learning (ML) models can utilize diverse data to enhance Earth system modeling, they often lack process understanding and adaptability to new conditions. We have developed surrogate modeling and generative ML methods that efficiently calibrate models using observations and quantify predictive uncertainties. Our approach, applied to ELM calibration, significantly outperformed the Markov Chain Monte Carlo method in predicting NEE with 30 times less computing time. We also introduced physics-informed, explainable ML methods with uncertainty quantification to enhance model trustworthiness and predictive understanding under new conditions. These methods have successfully simulated NEE and streamflow, capturing extreme water events and identifying key drivers for informed, climate-resilient decision-making and model development. Additionally, we are advancing AI foundation models to act as a digital Earth, training on extensive climate simulation and observation data for diverse spatiotemporal tasks from regional to global scales and subseasonal to multidecadal periods. Preliminary results from our billion-parameter climate foundation model demonstrate strong scalability on DOE’s Frontier supercomputers and robust prediction of key climate variables.