Virtual X-Ray Image Stack for Machine Learning Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional 3D simulations are resource-intensive and time-consuming, and existing methods for creating training sets for machine learning are limited by the complexity of 3D structures, making it difficult to efficiently model and analyze product development processes.
Innovation Solution
A method is introduced to generate a Virtual X-Ray Image Stack (VXRI stack) corresponding to a 3D structure, which includes multiple layers represented as 2D value matrices, allowing for the creation of a training set for machine learning algorithms to predict simulation results without the need for extensive 3D simulation data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional 3D simulations are used to model products and perform required simulations, then accurate performance attributes can be obtained, but significant time and computing resources are required
Solution Approach 1:
The patent creates virtual x-ray images as simplified copies of the complex 3D structure, capturing essential geometric features in a compressed 2D representation. This copying approach maintains the necessary information for simulation while dramatically reducing computational complexity and time requirements.
Solution Approach 2:
The patent extracts key geometric features from the full 3D structure by projecting them through virtual x-ray images. This extraction process isolates the most relevant structural information needed for simulation while discarding redundant data, thereby reducing computation time while preserving accuracy.
2Loss of information
If voxel representations are used to represent 3D topological structure, then complete structural information is captured, but the number of voxels becomes significant enough to make the process unmanageable
Solution Approach 1:
The patent transforms the 3D voxel data into 2D virtual x-ray images by projecting the three-dimensional structure along specific directions. This dimensionality reduction converts an unmanageable 3D dataset into compact 2D representations that are easier to process while retaining essential structural information.
Solution Approach 2:
The patent creates simplified 2D image copies of the 3D structure that capture the necessary geometric information without requiring the full voxel resolution. These image representations serve as efficient surrogates for the complete 3D data.
3Ease of manufacture
If parameters defining specific design features are used as training data, then the training set can be created, but the approach is strictly dependent on the underlying topology of the 3D structure which can be highly complex and very different among products
Solution Approach 1:
The patent creates a universal training data representation using virtual x-ray images that can be applied across different product topologies. This approach generates multi-functional training data that works for various structural configurations without requiring topology-specific processing, thereby improving versatility.
Solution Approach 2:
The patent transforms the training data from topology-dependent parameters into topology-independent image-based features. By changing the representation parameters from geometric measurements to pixel intensity values in virtual x-ray images, the system achieves adaptability across different product topologies.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach reduces the time and resources required for product development by enabling the prediction of 3D simulation results, allowing for earlier error detection and reducing the number of simulations needed, while being independent of the 3D structure modeling method.
Implementation Method 1
generating the image stack corresponding to the 3D structure includes projecting a ray through each grid point on a grid plane to the 3D structure
Implementation Method 2
integrating a weighted material property along a ray originating from an origin outside of the 3D structure through a grid point on an grid plane, from a predefined starting point to a predefined end point, to determine an integration result
Implementation Method 3
providing the training set to the machine learning algorithm to identify correlations between the image stack and the simulation results
Data Source
AI summary
Examples described herein relate to apparatuses and methods for determining a training set for an AI of a computerized simulation platform. An image stack corresponding to a 3D structure includes a plurality of layers generated based on the 3D structure. A model of the 3D structure is simulated to determine simulation results. The training set for a machine learning algorithm of the AI includes the image stack and the simulation results. The AI is trained using the machine learning algorithm based on the training set.