Neural network-based surrogate model in postprocessing of topology optimized structures.

Area of Science:

• Engineering
• Computational Mechanics
• Artificial Intelligence

Background:

• Topology optimization generates efficient structures but requires postprocessing for manufacturability.
• Converting topology-optimized designs to CAD smooths boundaries, altering stress distributions and requiring reconciliation.
• Finite Element Method (FEM) simulations are computationally intensive for iterative design refinement.

Purpose of the Study:

• To develop an accurate and efficient surrogate model for postprocessing topology-optimized structures.
• To minimize the computational expense associated with reconciling stress values between topology optimization and CAD models.
• To enable fine-tuning of geometry parameters for multiple stress performance metrics.

Main Methods:

• A feedforward deep artificial neural network (DANN) was designed with architecture parameters optimized for specific stress outputs.
• The DANN was trained using data generated from Design of Experiments (DoE) models, linking geometry dimensions to stress under various loads.
• A surrogate model was constructed using the trained DANN to predict stress performance metrics.

Main Results:

• The DANN-based surrogate model accurately predicted highly nonlinear stresses under combined loading conditions.
• Von Mises stress predictions achieved within 10% accuracy, and axial force sensor stress predictions within 2% accuracy.
• The method significantly reduced the number of required FEM computations for optimized post-processed designs.

Conclusions:

• The proposed DANN-based surrogate modeling approach is effective for postprocessing topology-optimized structures.
• This method offers a computationally efficient alternative to traditional iterative FEM-based postprocessing.
• The technique is validated by its successful application in postprocessing a wind tunnel balance design.