<p>The chemical composition of groundwater is often controlled by the mixing of water from multiple sources. Inferring both the number and the composition of these sources (end-members) from hydrochemical data remains challenging, due to the high dimensionality of the data and the widespread reliance on two-dimensional graphical interpretations. This paper therefore proposes a probabilistic framework for the robust estimation and detection of source composition in multidimensional chemical spaces. The unknown sources are modelled as a realisation of an interaction Gibbs point process, whose probability density incorporates expert knowledge typically employed in graphical mixing analysis. This includes mass conservation, bounded source space, a limited number of sources, and source distinctness. Source configurations are estimated by maximising the proposed probability density using a simulated annealing algorithm based on Metropolis–Hastings dynamics. The variability observed across optimal configurations generated at a very low temperature is analysed in order to characterise the operational variability and robustness of the detected sources under the imposed constraints, rather than as formal Bayesian posterior uncertainty. First, the method is calibrated using synthetic data, and then it is applied to two real hydrochemical datasets from geothermal and ore-forming hydrothermal systems. The results demonstrate that the proposed approach provides consistent and interpretable source estimates while offering a systematic way to integrate expert knowledge and evaluate the reliability of the inferred source configurations.</p>