Risk and Uncertainty Analysis of Error Propagation in Process Design and Simulation—Approaches and Modeling Results | AIChE

Risk and Uncertainty Analysis of Error Propagation in Process Design and Simulation—Approaches and Modeling Results

Authors 

Vasquez, V. R. - Presenter, University of Nevada, Reno
Risk and uncertainty analysis of error propagation in process design and simulation play an important role in the determination of error bounds of model predictions, design of safety factors, and risk estimation among other. Chemical process design and simulation rely heavily on computer-aided methodologies and thermophysical properties. The latter are commonly estimated through the use of empirical models, correlations, and theoretical models whose parameters are usually estimated from experimental data. In this presentation, we discuss how the issues of random and systematic errors in thermodynamic models affect their predictions and the impact of these on design and simulation with the analysis of a variety of situations and examples. Using Monte Carlo-based methods, we show that there is significant uncertainty in the design and operations of chemical processes when one considers the seemingly small uncertainty in data for vapor-liquid and liquid-liquid equilibria. A common approach in the chemical processing industries is to apply safety factors in design to account for the effects of uncertainties. However, Monte Carlo uncertainty analysis can be used to quantify the uncertainty and lead to rationalization of the safety factor approach. For example, sensitivity analysis of binary interaction parameters typical of models such as NRTL and UNIQUAC show that small variations in these translate into significant differences in performance evaluation of unit operations such as distillation and liquid-liquid extraction. The development of thermodynamic models from experimental data requires the regression of model parameters, which we use as input to the Monte Carlo simulations. However, the classical application of maximum likelihood methods for regression assumes that the model is inherently more accurate than the data. For phase equilibrium data, often the inverse is true. We show, with examples, the importance of separating the effects of different types of uncertainty such as those caused by model uncertainty and those by experimental uncertainty. Monte Carlo methods show to be a valuable tool in developing strategies to make this separation. If time allows, we will briefly discuss current initiatives in our research group towards developing methods for real-time risk assessment of patient safety in clinical processes using machine learning methods and big data; ideas that can be easily extended to many operations of the CPI.