Self-directed online machine learning for topology optimization.

Area of Science:

• Computational Engineering
• Artificial Intelligence
• Optimization Methods

Background:

• Topology optimization is crucial for material distribution in design.
• Traditional methods face computational challenges with many design variables.
• Finite Element Method (FEM) calculations are computationally intensive for large problems.

Purpose of the Study:

• To develop a computationally efficient method for topology optimization.
• To integrate Deep Neural Networks (DNNs) with Finite Element Method (FEM) for faster optimization.
• To enable solving large, multi-dimensional optimization problems.

Main Methods:

• Introduced Self-directed Online Learning Optimization (SOLO) integrating DNNs and FEM.
• DNN learns to approximate the objective function of design variables.
• Dynamically generated training data based on DNN predictions for adaptive learning.

Main Results:

• SOLO reduced computational time by 2-5 orders of magnitude compared to heuristic methods.
• Achieved convergence to the true global optimum through iterative DNN adaptation.
• Outperformed state-of-the-art algorithms in tested compliance minimization, fluid-structure, heat transfer, and truss optimization problems.

Conclusions:

• SOLO offers a highly efficient approach for complex topology optimization.
• The integration of DNNs with FEM significantly accelerates the design process.
• This method opens possibilities for solving previously intractable large-scale optimization challenges.