This
work shows a new approach to the parameter-fitting problem
useful in the solutions thermodynamic field, providing a more objective
framework to obtain better empirical/semiempirical models for chemical
engineering applications. A model based on the excess Gibbs energy
function g
E is used to represent the behavior
of real solutions, together with its first derivative h
E, using a combined modeling under the paradigm of multiobjective
optimization. The problem is formulated as an MINLP methodology to
simultaneously consider two aspects: the model complexity and the
best parametrization to prevent the overfitting, controlling the trade-off
between them by applying the Akaike Information Criterion to g
E residuals. Two different solvers, one deterministic
(SBB/CONOPT) and another evolutionary (GA), are used, and their ability
to solve the problem is analyzed. The designed methodology is applied
to three highlighted VLE cases in chemical engineering, and the results
obtained show the ability of the method to get the best model in each
case. The proposed methodology proved useful for modulating the number
of parameters considering the imposed requirements, which decrease
as the accuracy requirements for h
E are
relaxed. The efficient-fronts obtained show a small
trade-off region, noting that the proposed framework provides the
simplest models with the minimum completeness uncertainty.