Molecular Reconstruction of Heavy Petroleum Fractions: A Bayesian Approach

We developed molecular stochastic reconstruction algorithms from a Bayesian perspective. The parameter estimation of stochastic reconstruction models is a challenge due to their intractable likelihood functions. The use of the Bayesian optimization framework for likelihood-free inference provides us with a natural method for uncertainty propagation and model regularization, which reduces overfitting issues. Using only 9 parameters, the model was able to represent the properties of all vacuum residues studied in this work, both for the properties provided to the model, such as molecular weight, elemental analysis, SARA fractions, simplified nuclear magnetic resonance, and simulated distillation and for the properties not provided, such as full magnetic resonance. The posterior distributions of the parameters and properties showed the prediction uncertainties of the model with the credible intervals containing both constrained and unconstrained observed properties, thereby confirming its robustness. Besides, the posterior predictive mean was shown to be a good estimator for the observed properties.