A set-based approach to dynamic system design using physics informed neural network

Abstract

In the early stage of dynamic system development which has a multi-disciplinary and hierarchical structure, system requirements need to be cascaded down to target values of each component so that engineers can collaborate efficiently and concurrently. The purpose of this paper is to propose a novel set-based concurrent engineering method for a dynamic system by using machine learning. In the practice of the target setting study for concurrent engineering, both hierarchical simulations between system and component level and a solution to solve inverse problems using the simulation are required. The proposed method composes of two machine learning methods that satisfy these requirements. The first one is physics-informed long short-term memory (PI-LSTM) which enhances the mechanical modeling of component behavior. By applying the proposed PI-LSTM to where mechanical modeling is difficult, the adaptive range of mechanical modeling can be expanded. The PI-LSTM surrogate the dynamic behavior of the component model and can be used inside the system-level simulation. The other one is Bayesian active learning (BAL) applied to inverse problems to solve feasible regions where all system requirements are satisfied. In the proposed BAL, Gaussian process models are trained from the system-level simulation, and an acquisition function is evaluated and maximized to generate new sampling candidates. The set-based design using BAL has an advantage in the decoupling ability of design problems because feasible regions of each discipline sub-problem can be studied concurrently. To show the effectiveness of the proposed method, a numerical example of a vehicle design problem which has a hierarchical structure is demonstrated.

Keywords

• gaussian process
• bayesian active learning
• physics-informed neural network
• long short-term memory (lstm)
• design optimization
• surrogate modeling