Artificial Intelligence (AI) assumes a pivotal role in Earth science, leveraging deep learning’s predictive capabilities. Despite its prevalence, the impact of AI on scientific discovery remains uncertain. In Earth sciences, the emphasis extends beyond mere accuracy, striving for groundbreaking discoveries with distinct physical properties essential for driving advancements through thorough analysis. Here, we introduce a novel knowledge-guided deep symbolic regression model (KG-DSR) incorporating prior knowledge of physical process interactions into the network. Using KG-DSR, we successfully derived the Penman-Monteith (PM) equation and generated a novel surface resistance parameterization. This new parameterization, grounded in fundamental cognitive principles, surpasses the conventional theory currently accepted in surface resistance parameterization. Importantly, the explicit physical processes generated by AI can generalize to future climate scenarios beyond the training data. Our results emphasize the role of AI in unraveling process intricacies and ushering in a new paradigm in tasks related to “AI for Land Surface Modeling.”

人工智能 （AI） 利用深度学习的预测功能，在地球科学中发挥着关键作用。尽管人工智能很普遍，但它对科学发现的影响仍然不确定。在地球科学中，重点不仅限于准确性，而是努力获得具有独特物理特性的突破性发现，这对于通过彻底分析推动进步至关重要。在这里，我们引入了一种新的知识导向的深度符号回归模型 （KG-DSR），该模型将物理过程交互的先验知识整合到网络中。使用 KG-DSR，我们成功推导出了 Penman-Monteith （PM） 方程并生成了一种新的表面电阻参数化。这种基于基本认知原理的新参数化超越了目前接受的表面电阻参数化的传统理论。重要的是，AI 生成的显式物理过程可以推广到训练数据之外的未来气候情景。我们的结果强调了 AI 在解开复杂的过程复杂性和为与 “AI for Land Surface Modeling” 相关的任务引入新范式方面的作用。