General circulation models (GCMs) are essential for weather and climate prediction. They use physics-based simulations to model large-scale dynamics and small-scale processes. Recently, machine-learning models have matched or exceeded GCMs in weather forecasting accuracy, but struggled with long-term stability and ensemble forecasts. In this Nature paper , co-led by Janni Yuval, a new model, NeuralGCM, integrates machine learning with a differentiable solver for atmospheric dynamics. It performs as well as top machine-learning and physics-based methods for short-term forecasts and can track climate metrics accurately for decades with prescribed sea surface temperature. NeuralGCM offers significant computational savings and demonstrates that deep learning can enhance traditional GCMs in predicting the Earth system. Griffin Mooers also contributed to the research.