Neural Network Geometry Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for optimizing the geometry of physical structures are computationally expensive and unable to effectively utilize prior knowledge, requiring extensive computation times.
Innovation Solution
A neural network is trained using pictorial representations of physical structures and associated performance metrics, allowing it to infer performance metric values efficiently and reduce computational complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional optimization algorithms (FEM, genetic algorithms) are used to optimize geometry, then optimization accuracy can be achieved, but computation time becomes excessively long (weeks or months)
Solution Approach 1:
The patent applies preliminary action by pre-training a neural network on a dataset of geometric configurations and their corresponding performance metrics before the actual optimization process. This pre-computed knowledge is stored in the neural network's weights and biases, allowing the system to rapidly evaluate new geometric variants during optimization without performing full FEM simulations, thus dramatically reducing computation time while maintaining optimization accuracy.
Solution Approach 2:
The patent uses copying by creating a neural network model that replicates the input-output behavior of expensive FEM simulations. Instead of repeatedly executing the original FEM simulation process, the trained neural network serves as a computationally efficient copy that approximates simulation results, enabling rapid evaluation of geometric variants during optimization iterations.
2Adaptability or versatility
If traditional optimization algorithms are used, then optimization can be performed, but the system cannot effectively utilize prior knowledge from past optimizations
Solution Approach 1:
The patent implements preliminary action by collecting and processing prior optimization results into a training dataset before deploying the optimization system. The neural network is pre-trained on this historical data, embedding prior knowledge about geometric-performance relationships into the model's parameters. This allows the system to leverage lessons learned from past optimizations to guide current optimization efforts more effectively.
Solution Approach 2:
The patent applies feedback by using the neural network to rapidly evaluate geometric configurations during optimization, providing immediate performance feedback to the optimization algorithm. This closed-loop feedback mechanism allows the system to continuously learn and adapt, with each iteration informing the next, thereby effectively utilizing prior knowledge to accelerate the optimization process.
Data Source
Figure 1
Figure 2A~2C
Figure 3
AI summary
Systems and computer-implemented methods are provided for respectively optimizing a geometry of a physical structure with respect to at least one performance metric, and for training a neural network for use in optimizing the geometry of the physical structure with respect to the performance metric(s). The neural network may be trained to incorporate prior knowledge, e.g., from past optimizations and/or from measurements of physical embodiments of the physical structure, and may be used during the optimization of the geometry of a same type of physical structure to infer values of the performance metric(s), thereby reducing the computational complexity of the optimization compared to having to compute such performance metric values via FEM or similar methods.