An accurate and efficient procedure for estimating process design reliability

Safety factors are an expensive, but necessary part of any chemical engineering process design. Unfortunately, safety factor selection procedures are often little more than approximate rules of thumb based on experience. Increasing economic pressures and the risk of failure demand a rigorous procedure to size safety factors for maximum benefit at minimum cost. This requires statistical descriptions of the uncertainties and a method for calculating design reliability (odds of success).;Hybrid distributions (possibility and/or probability distributions) are the best (albeit undeveloped) tools for describing uncertainty. However, calculation of reliability is prohibitively expensive, requiring thousands of process simulations for adequate results. This research presents a new calculation method that is orders of magnitude better than previous methods. It provides accurate and precise reliability estimates for just tens or hundreds of process simulations. This includes essentially free and unlimited sensitivity analyses.;The key is to mathematically approximate the constraint boundary (border separating the regions of success and failure) independent of the statistics. This is done by finding and geometrically interpolating between points on the boundary. The statistical calculations can then be rapidly performed using the mathematical boundary-approximation. Process simulation effort is spent finding the boundary points instead.;While applicable to hybrid distributions the procedure was developed and tested with probability distributions only. The test problem was binary and ternary distillation with feed flowrate, tray efficiency, and thermodynamic database uncertainties. The procedure was shown to be orders of magnitude faster than Conventional Monte Carlo.;Computational requirements increase with the number of uncertain parameters, but are reasonable and remain much lower than for Conventional Monte Carlo. The reliability estimate may be slightly inaccurate--usually within 0.5% and always within 1%-2%--however, this error can be estimated and compensated for with some additional effort.;Distillation control variables were eliminated by implicit substitution. Other process operations may not allow this simplification.;Future work should extend the procedure to hybrid statistics, provide for explicit handling of control variables when necessary, and optimize the procedure. Most importantly, better statistical descriptions of chemical engineering process design uncertainties are required.