Résumé : This work proposes a Bayesian inference framework for estimating the low-dimensional dynamical operator of a reactive system. The framework can be employed both for non-intrusive reduced-order modeling and for Galerkin closure. In the first case, the choice of the prior is used to regularize the regression problem, and the optimal operator is inferred from snapshots of the system. Moreover, the prior is parametrized to ensure local and global stability of the inferred operator. In the second case, the Galerkin operator is given as a prior to the inference model, which improves the operator based on the collected data. To obtain a reduced-order representation of the dynamical system, we employ the proper orthogonal decomposition. Bayesian inference is carried out using a stochastic variational approach, enabling the use of non-Gaussian priors and likelihoods, thereby yielding complex posterior distributions. The proposed methodology is first applied to the Burgers’ equation, where we employ Bayesian inference both for non-intrusive reduced-order modeling and for Galerkin closure. Then, the same methodology is applied to the numerical simulation of a pulsated laminar flame. The results show that the proposed technique is capable of estimating the low-dimensional dynamical operator with good accuracy, resulting in a coefficient of determination between the observations and the forecast higher than 0.9 for all thermo-chemical features. The forecast uncertainty is estimated using an ensemble method, and it is crucial to assess the reliability of the prediction. Overall, this work advances the development of dynamical reduced-order models, enabling the real-time forecasting of combustion systems. Novelty and significance statement: The novelty of this work is the development of a Bayesian inference framework for estimating the low-dimensional dynamics of a reactive system, parametrized in a way that guarantees the local and global stability of the inferred operator. This framework can be applied for performing Galerkin closure and for non-intrusive reduced-order modeling. The inference is performed using a stochastic variational approach, which enables the estimation of complex posterior distributions arising from non-Gaussian priors. This work is significant because it introduces a unified probabilistic framework capable of performing both Galerkin closure and non-intrusive reduced-order modeling. The use of stochastic variational inference enables the parametrization that ensure a locally and globally stable operator. Furthermore, the Bayesian formulation provides two additional key benefits: it regularizes the regression problem and quantifies the uncertainty associated with the inferred operator. Overall, this work advances the development of dynamical reduced-order models of combustion systems.