Neural Network Component Design for Simulation Time Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current geometry modeling techniques, such as BRep, face limitations in controlling design space exploration, simulation time, reusing existing data, and addressing geometric and topological errors, which hinder efficient design and simulation processes.
Innovation Solution
The use of neural networks to generate component designs by obtaining datasets of component designs and performance values, categorizing them based on performance parameters, creating a training dataset, training an artificial neural network, and using it to generate new component designs based on specified performance criteria.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If BRep geometry modeling is used, then design representation is achieved, but design space exploration control is limited
Solution Approach 1:
The patent replaces traditional BRep mechanical geometry modeling with a neural network-based system that uses voxel representations and learned geometry manipulations, enabling more flexible design space exploration through data-driven approaches rather than rigid mathematical representations
Solution Approach 2:
The system transforms geometry representation from fixed BRep parameters to flexible voxel-based representations with learnable transformation parameters, allowing continuous exploration of design space through parameter variations that the neural network has learned to map to valid geometries
2Measurement precision
If Fine Element_analysis or CFD simulations are performed, then accurate performance prediction is achieved, but simulation time increases
Solution Approach 1:
The patent performs preliminary training of neural networks on simulation data to create surrogate models that can predict performance metrics instantly, eliminating the need to run full FEA or CFD simulations during design exploration while maintaining acceptable accuracy through the learned mappings
Solution Approach 2:
The system creates simplified voxel-based representations that copy the essential geometric features needed for performance prediction, allowing rapid evaluation without requiring detailed meshed models necessary for high-fidelity simulations
3Productivity
If existing simulation data is reused, then design efficiency is improved, but data integration difficulty increases
Solution Approach 1:
The patent merges existing simulation data with new design evaluations within a unified neural network training framework, combining datasets from different sources and types into a single learnable representation that leverages all available information without requiring complex separate integration processes
Data Source
AI summary
A method for generating new designs for a component using an artificial neural network, comprising: supplying a dataset of component designs represented as voxels; supplying a dataset of performance values, each performance value associated with either a respective component design or a voxel of a component design; categorising, based on the associated performance values, the component designs into performance categories according to one or more performance parameters; creating a training dataset by combining the performance categories for each component design, the dataset of performance values and the dataset of component designs; training an artificial neural network using the training dataset to produce a trained neural network; using the trained neural network to generate a new component design based on specified performance criteria.