Efficient multilayer coating design with elements of machine learning

Modern design tools allow solving a wide set of the most challenging multilayer design problems, but they require a deep knowledge of synthesis methods, and good expertise in applying various design tricks to obtain a good solution. During the last two decades computational performance of modern computers allows one to consider more computationally demanding approaches, including those ones based on deep search methods. Deep search approaches require enumeration of all possibilities on each iteration of a method. The number of different possibilities (layer insertion locations for the needle optimization, layer boundaries for the gradual evolution iteration, layer number for the design cleaner, etc.) is growing with the complexity of the coating. Therefore, the computational time of any deep search method considering all possibilities is also growing dramatically. To mitigate this challenge, we propose to use self-adapting algorithms accumulating information on main computational metrics (the numbers of required iterations, the evolution of merit functions versus iteration, the behavior of the merit function gradient norm). Accumulated information can be efficiently used to complete the computations of promising variants and to interrupt iterations of unfavorable variants very early. Proposed self-adapting and self-learning method serves as a superstructure upon well-known multilayer optimization methods, such as needle optimization, gradual evolution, design cleaner, etc. A huge advantage of the new self-learning method is that it does not require any careful fine tuning during computations. New method can run in completely automatic mode and nevertheless provide high quality solutions of challenging design problems.