OPTIMAL MOLECULAR DESIGN UNDER PROPERTY PREDICTION UNCERTAINTY

A systematic method that quantitatively assesses property prediction u ncertainty (imprecision) on optimal molecular design problems is intro duced. Property - structure relations are described with specific nonl inear functionalities based on group contribution methods. Property pr ediction uncertainty is explicitly quantified by using multivariate pr obability density distributions to model the likelihood of different r ealizations of the group contribution parameters. Assuming stability o f these probability distributions, a novel approach is introduced for transforming the original nonlinear stochastic formulation into a dete rministic MINLP problem with linear binary and convex continuous parts with separability. The resulting convex MINLP formulation is solved t o global optimality for molecular design problems involving many uncer tain group contribution parameters. Results indicate the computational tractability of the method and the profound effect that property pred iction uncertainty may have in optimal molecular design. Specifically, trade-off curves between performance objectives, probabilities of mee ting the objectives, and chances of satisfying design specifications o ffer a concise and systematic way to guide optimal molecular design in the face of property prediction uncertainty.