Development of an Evolutionary Deep Neural Net for Materials Research

Abstract

Modeling and optimization in many materials related problems routinely involve noisy, non-linear data from diverse sources. This novel algorithm, now tested on several problems, eliminates noise and extracts the meaningful trends from such data, using some multi-objective evolutionary algorithms, instead of the existing training methods. Some small neural nets with flexible topology and architecture are fed with random subsets of the problem variables, ensuring that each variable is used at least once. They evolve through a tradeoff between two conflicting requirements that they should be of maximum accuracy and at the same time of minimum complexity, defined through the number of parameters used. Mathematically, this leads to a Pareto-optimal problem, and the evolutionary algorithms that are used to train them are geared to handle that. These subnets are then assembled using a number of hidden layers; a linear least square algorithm is used for the optimization of the associated weights. Some applications in the metallurgical and materials domain are also discussed.