Extreme hot events led to catastrophic consequences on public health, economy, and crop losses over North America. While a great advance has been made in applying machine learning (ML) to weather forecast, whether ML is a winner in seasonal climate prediction of extreme hot events over North America remains unexplored. Here, by revealing the spatio-temporal characteristics of the leading modes of extreme hot days (EHDs) over the western North America (WNA), we set up a precursory predictor library for each of the leading EHDs modes and construct ML-based prediction models based on the library. Whereas the ML-based models exhibit nearly perfect cross-validation skills during the training period, the performance of the models during an independent prediction period is far from satisfactory. In contrast, a physics-based empirical (PE) model using six physically meaningful predictors shows better performance of prediction than the ML models. In particular, the PE model is able to predict the abnormal EHDs over WNA in 2021, whereas all the ML-based models fail.

北美热浪,机器学习,季节预测,物理机制