Reactive transport groundwater modelling is complex and computationally expensive, particularly in the risk-based decision-making framework that requires predictive uncertainty to be quantified. This limits its application in practical engineering solutions for groundwater contamination management and remediation at site/regional scale. Machine Learning (ML), in the form of a versatile data-driven surrogate model, has been tested recently to tackle the huge computational burden for regional groundwater flow modelling. However, its application in more complex numerical reactive transport models is rarely reported. Highly non-linear reactive transport models pose extra difficulty for emulator training. In the present study, three different ML models (Genetic Programming, Gaussian Process and Deep Neural Networks) are tested to substitute the process-based numerical model. Three types of predictions are emulated: (i) the contaminant concentration at a specific location and time (point prediction), (ii) the objective function used for model calibration and uncertainty analysis (lumped prediction) and (iii) the contaminant breakthrough at a specific location (time-series prediction). Although the three ML models can all potentially deliver a good performance, their requirements for the number of training samples and training time are different. The pros and cons of the different ML models will be discussed for each type of prediction. The numerical experiment is conducted based on a generalized reactive-transport model for onshore conventional gas fields in southeast South Australia where potential contamination pathways have been identified. Equipped with the efficient ML models, risk-based decision making can be supported by a reactive transport model. The emulators can also be integrated into operational water resource management platform to allow quick operational modelling.