The relationship linking leaf physiognomy and climate has long been used in palaeoclimatic reconstructions, but current models lose precision when worldwide data sets are considered because of the broader range of physiognomies that occur under the wider range of climate types represented. Our aim is to improve the predictive power of leaf physiognomy to yield climate signals, and here we explore the use of an algorithm based on the general regression neural network (GRNN), which we refer to as Climate Leaf Analysis with Neural Networks (CLANN). We then test our algorithm on Climate Leaf Analysis Multivariate Program (CLAMP) data sets and digital leaf physiognomy (DLP) data sets, and compare our results with those obtained from other computation methods. We explore the contribution of different physiognomic characters and test fossil sites from North America. The CLANN algorithm introduced here gives high predictive precision for all tested climatic parameters in both data sets. For the CLAMP data set neural network analysis improves the predictive capability as measured by R², to 0.86 for MAT on a worldwide basis, compared to 0.71 using the vector-based approach used in the standard analysis. Such a high resolution is attained due to the nonlinearity of the method, but at the cost of being susceptible to 'noise' in the calibration data. Tests show that the predictions are repeatable, and robust to information loss and applicable to fossil leaf data. The CLANN neural network algorithm used here confirms, and better resolves, the global leaf form–climate relationship, opening new approaches to palaeoclimatic reconstruction and understanding the evolution of complex leaf function.