Atmospheric turbulence phenomena are the main causes of injuries in civil air transport and due to climate change, the frequency and severity of turbulence is increasing [1]. There is no precise turbulence prediction method [2]. The state-of-the-art turbulence detection methods used on board commercial aircraft include pilot reports and online algorithms based on in situ eddy dissipation rate. They provide turbulence observations but not their predictions. Weather radar on the other hand only detects turbulence in wet air without any precise announcement about the timing. Equipped with a large number of sensors coming from different aircraft systems, the flight variables (multivariate time series generated by sensors) as well as their relationship may contain useful information indicating upcoming turbulence. Our approach aims at representing raw time series as functions which enable not only discovering the underlying function behind raw measurements but also implicitly removing data noise. Functional geometry features, which can capture the dynamic relation between variables, are deduced from the multidimensional path in functional representation. Based on the transformed geometry features, an outlier detection method is further deployed to detect specific behaviors indicating upcoming severe turbulence. Preliminary experimental results show that our approach reaches a 0.532 true positive rate while keeping a zero false positive rate, which meets the zero false alarm requirement for optimizing the passenger experience.