Engineering Design by Reinforcement Learning and Finite Element Methods
Nowadays research on aircraft design aim to reduce weight of airplane components, thus optimizing aircraft performance and reducing fuel consumption and operational costs. From this perspective novel materials and technologies are developed, but also advances in design methods and tools are put forward. Generative design is one of the approaches to automatically optimize component design. It uses evolutionary algorithms and topological optimization to generate original, unconventional and complex structures like novel bionic partition for Airbus A 320 cabin interiors, .
Alternatively, deep reinforcement learning has had great success in artificial intelligence applications. Among them, beating the champion of the game of Go in 2016, mastering many Atari games and optimizing the work of data centers. In my work, I have combined deep reinforcement learning (RL) and finite element analysis (FEA) for the purpose of automating structural design of components.
AI in general and deep reinforcement learning in particular is powerful approach in solving many nowadays problems in information technology, business, healthcare, and engineering. There is a myriad of applications for AI technologies that one can implement to make life easier. Structural engineering design is no exception. Designing a structure or a part of machinery is a very tiring process. One needs to make a lot of manual changes before resulting in the final design that satisfies structural loads. But this iterative process can be automated.
A typical approach to structural engineering design is finite element analysis. A number of authors have tried to combine FEA and machine learning [2–4]. For example,  have used deep-autoencoder to approximate the large deformations on a non-linear, muscle actuated beam. In , machine learning was used to predict the deformation of the breast tissues during the compression. However, little attention has been paid to using reinforcement leaning in assisting structural engineering design.
In my efforts, , I have tried to combine FEA and deep RL to assist an engineer in her design process. The results show that deep reinforcement learning in combination with finite element analysis can be used as automatic iterative process of structural engineering design.
In typical sitting, the finite element model represents an environment to which an agent applies actions and from which it gets observations and rewards (Figure 1). An agent uses neural network to decide on its actions. Actions change geometry of the structure, new geometry then subjected to FEA. Finite element analysis produces the state, which then is fed to neural network. And the process repeats itself. The agent gets rewards if it meets an optimization objective of minimizing target value (e.g. some nodal displacement in the structure) in each of the learning iterations. The end result of the modeling (after inference stage) is an optimized design of a structure. The inference stage is a usual greedy inference where an agent makes actions of altering the geometry based on observations only.
In the proposed design pipeline, an engineer provides initial geometry of a structure, sets loads and allowed actions to alter the geometry, specifies the optimization objective (e.g. minimize internal forces, minimize weight, maximize stiffness of a component), and starts training the model. After training, in inference stage, the engineer gets her final design thus showing that combination of FEA and RL might make structural engineering design automated.
To keep up with the project please see https://www.facebook.com/GigaTsk
 Generative Design: Advanced Design Optimization Processes for Aeronautical Application, S. Bagassi et al., ICAS 2016
 Machine Learning and Finite Element Method for Physical Systems Modeling. O.Kononenko, I.Kononenko, arXiv.org
 Towards Finite-Element Simulation Using Deep Learning. Francois Roewer-Despres, Najeeb Khan, Ian Stavness, CMBBE 2018
 A finite element-based machine learning approach for modeling the mechanical behavior of the breast tissues under compression in real-time. Martinez-Martinez F, et al. ComputBiot