Topology Optimization Using Neural Network Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing topology optimization methods rely on gradient information or heuristic criteria, which are not applicable for problems with severe nonlinearities or complex constraints, and are costly for high-dimensional optimization problems, especially when sensitivity information or heuristics are not available.
Innovation Solution
A dual optimization process using machine learning or stochastic optimization to generate update strategies for material redistribution in topology optimization, allowing for the application of non-gradient optimization strategies and reusing update strategies for similar optimization objectives with different boundary conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If gradient-based optimization methods are used for topology optimization, then optimization efficiency is improved, but applicability is limited to problems where sensitivity information is available
Solution Approach 1:
The patent replaces gradient-based optimization mechanisms with a neural network-based approach. The neural network is trained to directly map design variables to optimized topologies, substituting the traditional gradient descent mechanism. This allows the system to handle problems with severe nonlinearities and complex constraints where analytical gradients are unavailable, while maintaining optimization efficiency through the learned mapping.
Solution Approach 2:
The patent creates a neural network model that copies the optimization knowledge from training data. Instead of computing gradients for each new problem, the system uses the trained neural network to generate optimized topologies by copying patterns learned during training. This enables rapid optimization for new design problems without requiring sensitivity information, thus improving both efficiency and versatility.
2Manufacturing precision
If traditional topology optimization methods are applied to high-dimensional problems, then design space exploration is thorough, but computational cost increases significantly
Solution Approach 1:
The patent performs preliminary action by pre-training the neural network on a dataset of optimization problems. This training phase captures the essential optimization patterns and relationships in the high-dimensional design space. Once trained, the neural network can rapidly generate optimized topologies for new problems without requiring extensive computational resources, thus reducing the computational cost while maintaining design optimization quality.
Solution Approach 2:
The patent changes the parameter representation by using a neural network to directly predict optimized topologies from design variables. This transforms the traditional iterative optimization process into a direct parameter mapping problem. The neural network learns the complex relationships between design parameters and optimal topologies, enabling efficient high-dimensional optimization without the computational burden of traditional methods.
3Measurement precision
If problem-specific sensitivity analysis is performed for each optimization case, then optimization accuracy is improved, but method complexity and preparation time increase
Solution Approach 1:
The patent creates a universal neural network model that can handle multiple optimization problems with different objective functions and constraints. The single trained model serves multiple functions by predicting optimized topologies for various problem types without requiring problem-specific sensitivity analysis. This reduces method complexity while maintaining optimization accuracy through the generalized learning capability of the neural network.
Data Source
AI summary
In one aspect, a computer-assisted method for the optimization of the design of physical bodies, such as land, air and sea vehicles and robots and/or parts thereof, is provided comprising the steps of: representing the design to be optimized as a mesh, generating update signals to optimize the mesh representation, applying an optimization algorithm until a stop criterion has been reached, and outputting a signal representing the optimized design.