Reducing ecological impact is a major challenge for today’s industry, particularly the rail industry, which is one of the most energy-intensive industries. Indeed, this industry faces two paradoxical needs: on the first hand, it must decrease its energy consumption, meeting both an environmental goal and a financial objective, and on the other hand, it must not only maintain but increase the circulation of train, thus allowing a larger part of the population to use the most ecological means of land transport. Reducing consumption requires a prediction model. The formalization of this model is complex, particularly when driver control is taken into account. The complexity of the model must be chosen carefully. The SNCF has a very sophisticated model with a large number of parameters that need to be evaluated. In this work, we consider a train dynamics model simplified to a non-linear differential longitudinal dynamics equation coupled to a power balance [1]. In this model it is possible to distinguish two types of inputs: the model parameters which will be calibrated and the environment variables (wind, driver’s control, etc.) which will change from one journey to another. In order to calibrate the model, we rely on a priori knowledge from SNCF experts who provide information on the values of the model parameters. We also have a set of measurements (time, speed, power consumption, control) taken on the same train, for different journeys on different tracks. The used methodology is Bayesian calibration [2], which makes it possible to use the two types of information available to us while injecting model errors to take account of our imperfect knowledge of the system. These errors are parameterized by hyper-parameters that should also be calibrated. It is then necessary to formalize the experts’ knowledge mathematically in order to create our prior distributions, and then to write a likelihood function that will allow us to take into account both the measurements and the model errors. Another difficulty is the lack of knowledge about driver control as the control isn’t measured. The control linked to each measurement must be determined and appears in our problem as functional hyper-parameters. REFERENCES: [1] Julien Nespoulous, Christian Soize, Christine Funfschilling, and Guillaume Perrin. Optimisation of train speed to limit energy consumption. Vehicle System Dynamics, 60(10):3540–3557, October 2022. [2] Christian Soize. Uncertainty quantification: An accelerated course with advanced applications in computational engineering. Springer International Publishing, Cham, Switzerland, 2018.