Predicting the evolution of groundwater contamination is a major concern for society, in particular when investments are made to remediate the contamination. Groundwater reactive transport models are valuable tools to integrate the available measurements in a consistent framework, improving our understanding of the physical system and potentially revealing aspects of the system that were not considered before. In this thesis reactive transport models are developed for predicting groundwater contamination at a field site where a zero-valent iron permeable reactive barrier (PRB) is installed. In doing so, several general challenges with integrating data into reactive transport models are addressed. A first challenge is conceptual model uncertainty: available data should provide enough information for a correct conceptualization of the system. In chemical systems the concentration of each component is affected by several reactions and alternative reaction networks or descriptions of the reaction rates might provide similar model results. Moreover, the collection of new data might invalidate previous model conceptualizations and one might expect the conceptual model to be continuously updated as new information is acquired. Conceptual model uncertainties can have a large effect on reactive transport model predictions, yet they often are ignored. This thesis provides examples of quantification and reduction of conceptual uncertainty in reactive transport models. Quantification of conceptual model uncertainty was presented in the first part of this thesis, where a column scale experiment of a permeable reactive barrier is described. In the experiment, a substantial decline of the remediation performance over time was observed, mainly due to the development of a carbonate mineral coating around the iron particles. The column measurements were integrated in a multi-component reactive transport model, which subsequently provided predictions of long-term PRB efficiency under reduced flow conditions representative of the field site of interest. Different models of the deactivation process proposed in the literature were all able to reasonably well reproduce the column experiment measurements. The extrapolated long-term efficiency under different flow rates was however significantly different between the different models. These results highlight significant conceptual model uncertainties associated with extrapolating long-term PRB performance based on lab-scale column experiments. Nevertheless, possible improvements of the experimental design were suggested. Despite the large amount of data collected in the column experiment, model conceptual uncertainties could not be resolved clearly and simpler deactivation models might be justified. A simplified deactivation model was proposed, where the decline of the barrier reactivity is simply proportional to the effective groundwater velocity, in contrast to the complex geochemical model where detailed mineral precipitation reactions are accounted for. The main advantage of the simplified model over the geochemical model is that it does not require inorganic concentration measurements for the inference of its parameters. However, the simplified model failed in reproducing part of the column measurements as it does not take into account the feedbacks present in the true geochemical system, which might be important for predicting the long term barrier performance. Reduction of conceptual model uncertainty was demonstrated for a field-scale reactive transport model, where alternative descriptions of the groundwater recharge process provided similar simulations of the groundwater levels but different simulations of the contaminant plumes, with different estimates of the hydraulic conductivity fields. In this case, model conceptual uncertainties were partially solved by integrating simultaneously in the model groundwater heads, concentrations and direct estimation of the hydraulic conductivities. A second challenge concerns the data integration method. Since reactive transport models describe the evolution of multiple species in space and time, multiple data sets are used in parameter inference, for example measurements of dissolved concentrations of various chemical species. A common approach is to optimize parameters using an overall measure of fit, such as the weighted sum of squared residuals, where each data set is given a weight based on prior knowledge of the measurement and model errors. In practice, model errors are difficult to determine a priori and often they are much larger than measurements errors. Moreover, model errors induce correlations between residuals, which are typically ignored. A multivariate approach (MV) was proposed to integrate different data types, which accounts for specific correlations between residuals. The MV method allows integrating out the (co)variances (or weights) from the parameter posterior leading to an efficient estimation of model parameters, with no a priori assumptions about the magnitude of the errors. When applied to inference of the parameters of a column-scale reactive transport model, it was shown that accounting for residual correlation between species provides more accurate parameter estimation for high residual correlation levels, whereas its influence for predictive uncertainty is negligible. A limitation of the multivariate method is that it cannot be implemented using a full covariance matrix with a dimension equal to the number of all measurements. Instead, residuals must be grouped such that the number of groups does not exceed the number of measurements per group. A third challenge for field-scale models is the characterization of the spatial heterogeneity of soil properties from the measurements. Generally, little is known about heterogeneity and often soil properties are assumed constant within parts of the model domain or even in the entire model domain. However, heterogeneity has a considerable effect on the spreading of pollutants and must be accounted for to provide reliable predictions of contamination, increasing the number of model parameters to be estimated. Increasing the number of model parameters has two negative effects on parameter inference. First, it increases the computational effort. This is particularly relevant for reactive transport models characterized by long simulation times (ranging from several hours to days for field-scale applications). Second, a large number of model parameters can make the inverse problem ill-posed, where at least one of the criteria for well posedness (existence, uniqueness and stability) is not met. The reactive transport model developed for predicting the evolution of the contamination at the PRB site is affected by both of these two negative effects, since it accounts for spatial heterogeneity and several physical and chemical processes that increase the simulation time. The computational issue is solved using high performance computing. Ill-posedness is alleviated using a regularization procedure and a step-wise approach for integrating different data sets into the model. It is shown that the joint estimation of flow and transport parameters from head and concentration data improves the matching of the simulations to the measurements compared to the separate estimation of flow and transport parameters, in particular for the more mobile contaminants (cis-DCE and VC). Moreover, a more realistic estimation of the effective porosity and a reduction of the dispersion coefficients are obtained in the joint inversion. Additional measurements are required to validate the estimated recharge fractions and hydraulic conductivities. In the last part of the thesis, the inferred hydraulic and chemical parameters were used to predict future contaminant migration at the field site. These scenarios include all processes considered in the parameter inference, also accounting for the iron deactivation process due to mineral precipitation. The results indicate that the PRB is expected to remain effective at degrading in-situ groundwater contamination until at least 2035. However, existing contamination downstream of the reactive barrier is likely to persist beyond 2035, due to the small dilution effect of infiltrating rain and the small biodegradation rates at the site.