Accurate instantaneous vehicle emissions models are vital for evaluating the impacts of road transport on air pollution at high temporal and spatial resolution. In this study, we apply machine learning techniques to a dataset of 70 diesel vehicles tested in real-world driving conditions to: (i) cluster vehicles with similar emissions performance, and (ii) model instantaneous emissions. The application of dynamic time warping and clustering analysis by NOx emissions resulted in 17 clusters capturing 88% of trips in the dataset. We show that clustering effectively groups vehicles with similar emissions profiles, however no significant correlation between emissions and vehicle characteristics (i.e. engine size, vehicle weight) were found. For each cluster, we evaluate three instantaneous emissions models: a look-up table (LT) approach, a non-linear regression (NLR) model and a neural network multi-layer perceptron (MLP) model. The NLR model provides accurate instantaneous NOx predictions, on par with the MLP: relative errors in prediction of emission factors are below 20% for both models, average fractional biases are −0.01 (s.d. 0.02) and −0.0003 (s.d. 0.04), and average normalised mean squared errors are 0.25 (s.d. 0.14) and 0.29 (s.d. 0.16), for the NLR and MLP models respectively. However, neural networks are better able to deal with vehicles not belonging to a specific cluster. The new models that we present rely on simple inputs of vehicle speed and acceleration, which could be extracted from existing sources including traffic cameras and vehicle tracking devices, and can therefore be deployed immediately to enable fast and accurate prediction of vehicle NOx emissions. The speed and the ease of use of these new models make them an ideal operational tool for policy makers aiming to build emission inventories or evaluate emissions mitigation strategies.