For process simulation involving chemical reactions, thermodynamically accurate model parameters for chemical equilibrium are indispensable. We demonstrate that regression subject to the well-known chemical equilibrium constant entails violation of the sufficient conditions for thermodynamic stability. As a consequence, erroneous values of the equilibrium conversion are obtained upon use of the respective parameter values in, e.g., process simulation. In order to alleviate the problem, we propose a methodology for estimating parameters subject to necessary and sufficient criteria for thermodynamic stability. In particular, we propose an extension of Mitsos et al. (2009) to parameter estimation in reactive systems; an upper-level model-experiment mismatch is minimized subject to several lower-level minimization problems (LLP), which ensure the fulfilment of the reactive tangent plane criterion (RTPC) of Smith et al. (1993), along with other desired behavior of the Gibbs free energy. A fictitious single-phase dimerization reaction employing Margules and NRTL as Gibbs free energy models illustrates the thermodynamic guarantees provided by the proposed formulation, when the conventional ones fail. Our formulation conceptually allows any type of phase equilibrium, e.g., VLLE, as well as any type of activity coefficient model. The subproblems of the bilevel algorithm pose challenges to state-of-the-art commercial solvers. (C) 2018 Elsevier Ltd. All rights reserved.