Engineering Component Generation Through MLP-VAE 3D Field Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for storing and managing large volumes of 3D simulation data from engineering components, such as turbine blades, are inefficient and costly, leading to data deletion or loss of valuable information due to storage limitations.
Innovation Solution
A hybrid neural network model combining a variational autoencoder (VAE) and a multilayer perceptron (MLP) is used to compress and store 3D field data, allowing for efficient data storage and prediction of new design variations without the need for costly simulations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional storage technologies are used to store 3D simulation data, then data can be stored, but storage space requirements become enormous and costly
Solution Approach 1:
The patent applies a composite machine learning approach combining variational autoencoder (VAE) and multilayer perceptron (MLP) models to create a hybrid architecture that achieves superior compression ratios. The VAE component learns the underlying distribution of simulation data while the MLP component efficiently maps geometry parameters to this distribution, together achieving over 99% compression of CFD and FEA data.
Solution Approach 2:
The patent transforms the storage problem by changing the representation parameters of the data. Instead of storing raw 3D field data with millions of nodes, the system converts the data into a compressed latent space representation with only dozens of parameters, fundamentally changing how the data is stored and accessed.
2Volume of stationary object
If 3D simulation data is compressed using conventional methods, then storage space is reduced, but data quality and detail are lost
Solution Approach 1:
The patent replaces conventional mechanical compression algorithms with a machine learning-based compression system. The VAE-MLP model learns the intrinsic patterns and relationships in simulation data, enabling lossless or near-lossless compression that preserves all critical engineering details while achieving over 99% reduction in storage requirements.
Solution Approach 2:
The system changes the parameter representation from raw simulation data to a compressed latent space encoding that captures all essential information. The trained model transforms millions of data points into a compact set of parameters that can be stored efficiently while maintaining full data fidelity for future analysis.
3Volume of stationary object
If individual scalar values per simulation are stored, then storage requirements are reduced, but valuable 3D field data cannot be reused for future optimizations
Solution Approach 1:
The patent creates a compressed copy of the full 3D field data in latent space that retains all essential information. This compressed representation can be stored efficiently and then decoded or used directly for various future applications including design optimization, sensitivity analysis, and generating new design variations without requiring the original full-resolution data.
Solution Approach 2:
The compressed latent space representation serves multiple purposes: it can be used for storage efficiency, for generating new design variations, for performing sensitivity analyses, and for training other machine learning models. This single compressed format replaces multiple separate data storage needs.
4Quantity of substance
If machine learning models are trained to compress 3D field data, then compression ratios exceed 99%, but computational resources are required for training
Solution Approach 1:
The patent performs the computationally intensive training of the VAE-MLP model once during an initial setup phase. After this preliminary action, the trained model can be stored and reused indefinitely for compressing and generating simulation data without requiring repeated training, making the initial computational investment worthwhile for long-term data management.
Data Source
Figure 1~2(d)
Figure 3~5
Figure 6~6(b)
AI summary
Method of generating an engineering component, and method of saving 3D field data from simulations or measurements The invention relates to a method of storing 3D field data, the method comprising the following steps: S1 - Obtaining 3D field data related to a first group of specimens of a parameterizable engineering component, the 3D field data comprising geometry data; S2 - Defining a set of geometry parameters of the specimens; S3 - Training a variational auto encoder model (VAE) to compress the 3D field data to a latent space and restore it, and then splitting the VAE into an encoder model and a decoder model, wherein learned nodal weights of the encoder model and decoder model are set permanent for the further steps of the method; S4 - Connecting a multilayer perceptron network model (MLP) to an input layer of the decoder model of the VAE to form a Hybrid Multilayer Perceptron - Variational Autoencoder model (MLP-VAE) ; S5 - Training the MLP-VAE to map the set of geometry parameters to the 3D field data; and S6 - Storing the latent space in a storage medium.