This study implements deep learning techniques to estimate fuel burn of a jet aircraft. Current ground-based flight planning systems utilize aircraft type specific performance tables to determine fuel flows for given flight conditions and parameters such as altitude, mass and speed. These tables are corrected by a performance factor as the aircraft ages. Despite this update, planned fuel consumption may indeed not overlap with the actual one. In order to synchronize the base aircraft model with aircraft's actual performance, we propose using state-of-the-art deep learning algorithms for building data-driven models of fuel flows. Towards this goal, aircraft's on-board recorded trajectory and parameter data, namely Quick Access Recorder (QAR) data are utilized. The total dataset used within this study comprises of more than 1000 B777-300ER flights from a major European flag carrier airline. The deep neural network architecture is utilized for modeling the actual fuel flow specific to each aircraft and for each major flight mode (climb, cruise and descent namely). We have developed three neural network architectures (according to in-flight and ground based planning use cases) to present a tail-number specific correction factor to Base of Aircraft Data (BADA) models. First architecture involves a QAR data based black-box fuel flow model utilizing in-flight throttle data from all the engines. Comparison of this model with real flight data shows that precise estimation of fuel flow with mean errors lower than %0:1 can be achieved. The second architecture utilizes a physically consistent data regeneration of ffiT (delta thrust) using BADA formulation as to account for the ground planning phase where throttle information is not available. The third model involves a cascaded architecture which utilizes a neural network throttle estimator and the blackbox QAR fuel flow model for again the ground planning phase. Comparison of the latter models with real flight data shows that precise estimation of fuel flow with mean absolute errors lower than %0:7 can be achieved at all the flight modes. Initial tests reflect the fact that even better accuracy can be achieved for all models as the data set size increases. Finally, ground based planning fuel flow models are applied to actual flight plans generated by ground based systems. Total trip fuel comparisons show discrepancies up to %3:5 total fuel loading weight, which may result in potential fuel savings by decreasing the fuel load during take-off. Comparison of the planned and the estimated fuel boarding weights (following the actual filed flight plans) for 91 long-haul flights show that fuel burn savings of around 800[kg] per flight could have been achieved by the proposed methodology at the ground planning phase. For a typical operation of 100 long-haul flights per day, this represents yearly savings on the order of around 17 million USD at current jet fuel prices. This tail-number specific" performance modeling approach is projected to open considerable frontiers including the in-flight update of performance models through machine learning methods.