MIOpt: optimization framework for backward problems on the basis of the concept of materials integration

Abstract

In materials design, it is very difficult to accurately design forward problems owing to the variety of scales and phenomena to be considered and the increasing number of input and output variables. In addition, it is necessary to reduce the number of computationally expensive calculations to perform backward problem to obtain the optimal set of parameters for the input to a forward problem. In this paper, we describe the development of the MIOpt (MI Optimizer), a framework for backward problem. The process – structure – property – performance relationship (PSPP relation) of materials is modeled as forward problems in MInt, and constructed in terms of materials engineering concepts. This can be linked to backward analysis via MIOpt, which greatly improves the efficiency of trial – and – error materials design on a computer. In addition, two proprietary optimization algorithms are implemented, allowing the use of an extension of Bayesian optimization and an update algorithm for machine learning models, which reduces the actual time required for optimization. The benchmark problems were the minimum search problem for the analytical function and the heat input condition optimization problem for creep rupture time in welds. Efficient Bayesian optimization method was confirmed to be faster and more stable than the conventional sequential Bayesian optimization method.

Keywords

• materials integration
• backward problem
• bayesian optimization
• cubist model
• heat affected zone (haz)
• creep rupture time