Separating global horizontal irradiance measurements into direct and diffuse components has been vigorously discussed over the past half-century of solar radiation research leading to the creation of many models which attempt to compute these components with varying degrees of success. However, over the course of this discussion, nearly all studies have focused on hourly values, with no studies that have proposed a model for minute-level values of irradiance. As data-logging technologies have become much more prolific and their storage capabilities much larger, solar radiation monitoring sites are more commonly logging data at intervals much less than one hour, but no models exists that are designed to separate these measurements into direct and diffuse components. In Australia, the Australian Bureau of Meteorology and the Australian Solar Institute have compiled a dataset of tens of millions of one-minute global, direct and diffuse solar irradiance observations, comprising data from regions all around Australia. This dataset provides a unique opportunity to investigate the relationships between global irradiance and its direct and diffuse components at higher resolution than has previously been possible. Herein, the largest and most complete diffuse fraction model analysis yet undertaken for Australian solar radiation data, and. the first ever to focus on minute resolution data is reported. Nine of the most prominent diffuse fraction, or "separation", models are tested against minute resolution radiation data from three datasets. The first removed cloud enhancement events in accordance with practices undertaken by the majority of studies in the literature. The second retains these events in order to assess which model would be best suited for operational purposes. The third consisted of only clear sky observations, in order to assess the performance of diffuse fraction models under clear skies. Through the course of this study only the Perez model was found to perform satisfactorily for minute resolution data at sites in southeastern Australia. Three new diffuse models proposed in this study, one trained for each of the three datasets, were found to greatly exceed the performance of existing modeling techniques, with slight improvements over the Perez model. (c) 2015 Elsevier Ltd. All rights reserved.