Implementation of the multidirectional search algorithm on an automated chemistry workstation. A parallel yet adaptive approach for reaction optimization

Abstract

We have developed an experiment-planning module for applying a powerful new pattern search algorithm toward the problems of reaction investigation and optimization. The experiment planner works in conjunction with a closed-loop automated chemistry workstation equipped for parallel experimentation. The new algorithm, developed by Torczon for parallel computation of mathematical functions, achieves a focused yet parallel approach to finding regions of improved response. Like a full factorial design, the search space involves a regular grid of points. Like the Simplex algorithm, the multidirectional search (MDS) algorithm uses the movement of a simplex through a search space. However, with each movement all points except the single best are discarded, whereas the Simplex algorithm discards only the one worst point. Thus, in an n-dimensional space, the MDS algorithm projects n mandatory points at every cycle (beyond the initial). In addition, a larger number of exploratory points are identified by look-ahead projection of possible future simplices. Such exploratory points lie on multiple independent lines of search. The responses for the mandatory and exploratory points are acquired via parallel experimentation, with the latter points examined to the extent that the workstation has available capacity during the same schedule. The data from such exploratory points can be used in later cycles of experimentation, accelerating convergence on the region of optimal response. In the case of unlimited parallel experimentation capacity, all possible points in the space are projected, as in a full factorial design. The MDS algorithm thus adapts to the available parallel capacity of the workstation. The MDS planning module includes options for specifying initial points, stop criteria, and early-termination processes. Provisions are included for parallel scheduling of batches of experiments, convergence of the search, and movement at the boundaries of the search space. An MDS investigation can thus be implemented with global decision-making concerning movements through a search space, and local decision-making concerning termination of individual experiments. The MDS algorithm enables directed evolutionary searches in a parallel mode and is ideally suited for rapid optimization of chemical reactions using a microscale automated chemistry workstation.