Earth system models have been continously improved over the past decades, but systematic errors compared with observations and uncertainties in climate projections remain. This is due mainly to the imperfect representation of subgrid-scale or unknown processes. Here we propose a next-generation Earth system modelling approach with artificial intelligence that calls for accelerated models, machine-learning integration, systematic use of Earth observations and modernized infrastructures. The synergistic approach will allow faster and more accurate policy-relevant climate information delivery. We argue a multiscale approach is needed, making use of kilometre-scale climate models and improved coarser-resolution hybrid Earth system models that include essential Earth system processes and feedbacks yet are still fast enough to deliver large ensembles for better quantification of internal variability and extremes. Together, these can form a step change in the accuracy and utility of climate projections, meeting urgent mitigation and adaptation needs of society and ecosystems in a rapidly changing world.

过去几十年来，地球系统模型不断改进，但与观测结果相比的系统误差和气候预测的不确定性仍然存在。这主要是由于亚网格尺度或未知过程的不完美表示。在这里，我们提出了一种具有人工智能的下一代地球系统建模方法，需要加速模型、机器学习集成、地球观测的系统使用和现代化基础设施。这种协同方法将能够更快、更准确地提供与政策相关的气候信息。我们认为需要采取多尺度方法，利用千米尺度的气候模型和改进的粗分辨率混合地球系统模型，其中包括基本的地球系统过程和反馈，但仍然足够快以提供大型集合，以便更好地量化内部变异性和极端情况。总之，这些可以使气候预测的准确性和实用性发生阶跃变化，满足快速变化的世界中社会和生态系统的紧迫缓解和适应需求。