(116c) Guided Experimental Design for Modeling Reaction Rate Expressions

Sequential experimental design uses knowledge from previous experiments to guide what experiments to run next. In the literature, this goes by many names; optimal experimental design, adaptive experimental design, and goal-oriented experimental design. Guided or goal-oriented experimental design methods help achieve the target cheaply and quickly. Recently, there has been a surge in using Bayesian optimization (BO) for guided experimental design. BO uses non-parametric regression to model an objective, and sequentially guides the optimization by maximizing an acquisition function. At every iteration, BO determines the next evaluation point and tries to sample in areas where the uncertainty is high (for explorative purposes) and in areas close to where the current best is (for exploitive purposes).

In this work, the experimental objective is obtaining accurate and reliable models for reaction rate expressions. BO convergence rates will be evaluated for different cases. The effect of increasing the dimension of the input variables, level of noise, and number of hyperparameters of the kernel functions on the BO convergence rates will be studied. Furthermore, BO becomes computationally and statistically inefficient for high dimensional systems. Developing BO methods that work well in higher dimensions is thus of great practical and theoretical interest. A large body of literature has been devoted to mitigate the challenges of high dimensional BO problems. Several of which exploit the intrinsic lower dimensionality of the objective function, and perform BO in the lower dimensional space, followed by an inversion back to the high dimensional space. In this work, the problem of higher dimensions will be addressed for the reaction modeling application where a limited number of initial samples creates a unique dimensionality reduction problem.