oa Probing equation of state parameter fitting in parallel computers

The accurate design of chemical processes depends on the availability of models to predict the physical properties of the materials being processed. Thermodynamic properties such as enthalpies, entropies, and fugacities are particularly important in this context. Most of the models to evaluate them have adjustable parameters, fitted to give the best possible representation of the experimental data available for a given substance or mixture. Depending on how much information is available, this may entail the use of hundreds or thousands of data points. As several modern thermodynamic models have intricate mathematical expressions, especially equations of state, using so many data points to fit their parameters leads to substantial computational effort. This makes it difficult to run the parameter fitting problem from different initial estimates. The consequence is that this decreases the likelihood of finding the global minimum of the objective function used for parameter fitting. Despite the fact that current desktops and laptops are capable of parallel computations, little has been done to take advantage of their computational power for equation of state parameter fitting. The authors have recently developed procedures to that end, executed in different desktop and laptop computers, which provided speedups compatible with the number of processors available. One of the procedures is based on the conventional, sequential simplex minimization algorithm with a parallel evaluation of the objective function (SSPO approach). The other procedure is based on a modified, parallel version of the simplex minimization algorithm with a sequential evaluation of the objective function (PSSO approach). In this paper, we extend the evaluation of these procedures, executing them in the Suqoor supercomputer of Texas A&M University at Qatar, using single and multiple nodes. Because of numerical algorithm used, speedups in the PSSO approach are limited by the number of parameters to be fitted, which does not happen in the SSPO approach. On the other hand, the PSSO approach often ends at solutions with smaller objective functions, showing a greater tendency to escape local minima.