Convolution Modeling for 3D Print Shape Deviation Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing physics-based modeling frameworks for additive manufacturing (AM) are computationally intensive and do not provide a clear structure for machine learning, failing to account for process and path dependencies that cause inaccuracies in 3D shape accuracy due to swelling, shrinkage, and delamination during the manufacturing process.
Innovation Solution
A convolution modeling and learning framework is employed to predict geometric shape accuracy by generating representative transfer functions, decoupling geometric shape complexity from system response, and incorporating machine learning techniques to analyze interlayer interactions and error accumulations, using Fourier bases and Gaussian process regression to refine models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If physics-based modeling and simulation approaches are used to describe object formation from points to lines, lines to surfaces, and surfaces to 3-D shapes, then a detailed voxel-level description of the manufacturing process is achieved, but the computational complexity becomes excessively high and the structure is not suitable for machine learning
Solution Approach 1:
The patent segments the continuous physics-based modeling process into discrete convolution operations. By representing the manufacturing process as a series of layer-by-layer convolution operations (y = f * g), the complex continuous simulation is divided into manageable discrete steps that can be efficiently computed and learned by machine learning models, reducing computational complexity while maintaining prediction accuracy.
Solution Approach 2:
The patent replaces the traditional physics-based mechanical simulation system with a data-driven convolutional neural network system. Instead of solving complex physics equations for thermal fields, melt pool geometry, and cooling cycles, the system uses learned convolution kernels to directly predict shape deviations, substituting the mechanical physics simulation with an efficient neural network-based approach.
2Reliability
If traditional physics-based modeling is used, then detailed process description is achieved, but it fails to account for process and path dependencies that cause swelling, shrinkage, and delamination
Solution Approach 1:
The patent introduces dynamic adaptability through learned convolution kernels that can capture varying process dependencies. The convolution operation y = f * g allows the system to adaptively learn how previous layers (f) influence current layer deviations (y) through the kernel (g), enabling the model to account for dynamic process dependencies like swelling, shrinkage, and delamination that change throughout the manufacturing process.
Solution Approach 2:
The convolution framework inherently provides feedback mechanisms where the output of previous layer predictions influences subsequent predictions. The learned kernel g captures the feedback relationship between accumulated layer deviations and current shape accuracy, allowing the system to account for path dependencies by continuously incorporating information from previously manufactured layers into current predictions.
3Loss of information
If voxel-level physics-based simulation is performed, then comprehensive process description is achieved, but machine learning of AM data lacks a clear structure
Solution Approach 1:
The patent creates a universal convolution framework that serves multiple functions simultaneously: it preserves process information through the convolution operation, provides a clear structure for machine learning, and enables both prediction and interpretation of shape deviations. The unified model y = f * g can handle various additive manufacturing processes and geometries, making the framework universally applicable while maintaining computational efficiency and structural clarity.
Data Source
AI summary
A system, method, and computer-readable medium having machine instructions provides for predicting geometric shape accuracy of 3D printed products. Such a prediction may involve developing a model of an object and determining ways in which an actually-manufactured 3D object corresponding to the model differs in real life. These differences correspond to shape deviations. Shape deviations may be process dependent and/or path dependent, and different layers of the object as well as the manufacturing process to make the different layers may introduce shape deviations in layers of the object. By developing a transfer functions of the manufacturing process and associated interlayer effects of the layers and then appropriately offsetting inputs to the manufacturing process, the shape deviations may be ameliorated.