Abstract. Building climate models is a typical means of studying the dynamics of the climate system and assessing the impacts of climate change. However, model-related biases are common in existing climate models, such as the double ITCZ bias in most CMIP5 models. Recent studies suggest that biases showing distinct spatio-temporal characteristics may involve different mechanisms and sources in climate models. More dedicated studies on bias patterns is important not only for improving model performance, but also for helping modelers to better understand the climate system. In this paper, we focus on detecting spatio-temporal consistent bias patterns from climate model outputs. A spatio-temporal pattern is a bias pattern that is present in contiguous space with significant and coherent biases in continuous time periods. These patterns are ideal for revealing regional and seasonal characteristics of biases and worth further investigation by modelers. Due to the high computation cost, most of the existing analysis methods can only detect bias patterns that are either spatial consistent or temporal consistent, but not both at the same time. We proposed a bottom-up algorithm to tackle this problem. The proposed method first detects regions showing significant and consistent biases at each time slot and then merges them iteratively to form bias instances. The resulting bias instances are further clustered into different families to depict corresponding spatio-temporal consistent bias patterns. The experiments on both MIROC5 and FGOALS-g2 precipitation outputs show that the proposed approach can detect some important bias patterns that are consistent with previous studies and can produce other interesting findings. Modelers can adopt the proposed method as an exploratory tool to develop insights for bias correction and model improvement.