Layer-Wise AM Behavior Prediction Using Scan Maps and Prior-Layer State
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing additive manufacturing techniques, such as Laser Powder Bed Fusion (LPBF), face challenges with thermal distortion due to high cooling rates and temperature gradients, leading to thermal stress, which are not effectively addressed by current simulation methods.
Innovation Solution
A machine learning model is employed to predict the physical behavior of layers in additive manufacturing by using two-dimensional maps of laser scan properties and a state map of the prior layer, trained to accurately forecast thermal strain and other behaviors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If finite element analysis simulations are used to obtain thermal strains, then manufacturing precision can be improved, but productivity deteriorates due to long simulation times
Solution Approach 1:
The patent pre-processes laser scan instructions into 2D maps representing scan properties (power, speed, hatch spacing) before actual manufacturing. These pre-computed maps serve as inputs to the machine learning model, enabling rapid prediction of thermal strains without requiring time-consuming finite element simulations during the manufacturing process itself.
Solution Approach 2:
The patent replaces the traditional finite element analysis mechanical simulation system with a machine learning-based predictive system. The ML model is trained on simulation data to learn the complex thermal-mechanical relationships, then uses this learned knowledge to rapidly predict thermal strains without executing physical simulations, achieving both speed and accuracy.
2Manufacturing precision
If detailed simulations are performed for each layer, then manufacturing precision is improved, but loss of time increases significantly
Solution Approach 1:
The patent segments the 3D manufacturing process into individual 2D layers, with each layer processed independently. The laser scan instructions are divided into layer-specific 2D maps, and the ML model predicts thermal strains for each layer separately using only that layer's scan properties and the deformed state from the previous layer, avoiding the need for full 3D simulations.
Solution Approach 2:
The patent creates simplified 2D map representations (copies) of the laser scan instructions for each layer. These 2D maps capture the essential scan properties (power, speed, hatch spacing) in a compressed format that the ML model can process rapidly, replacing the need for detailed 3D simulation data while preserving the necessary information for accurate thermal strain prediction.
3Manufacturing precision
If comprehensive simulation models are used to account for thermal stress and distortion, then manufacturing precision improves, but device complexity increases
Solution Approach 1:
The patent transforms the complex simulation problem into a different parameter space by using 2D maps of scan properties (power, speed, hatch spacing) as inputs to the ML model. The model learns the complex thermal-mechanical relationships during training and then predicts thermal strains by processing these simplified 2D parameter representations, avoiding the need for complex simulation models during actual manufacturing.
Solution Approach 2:
The patent introduces 2D maps of laser scan properties as an intermediary representation between the laser instructions and the thermal strain prediction. These 2D maps serve as a simplified intermediate format that captures the essential scanning parameters in a structured way that the ML model can efficiently process, replacing the need for complex simulation models.
Data Source
AI summary
A three-dimensional model (i.e., MCAD model) for additively manufacturing a physical object with a laser is received in a computer system. The laser is controlled or driven by a set of instructions. The instructions specify scan properties for scanning each of multiple deposit layers during the additive manufacturing. Two-dimensional (2D) maps representing the scan properties for the deposit layer are generated. Each 2D map includes values in grid points representing the deposit layer. Each 2D map corresponds to a respective scan property for the deposit layer. A trained machine learning model is invoked to predict a physical behavior of a layer of the physical object based on the 2D maps as an input. The input further includes a state map showing a state of an immediately prior deposit layer. The layer of the physical object corresponds to the deposit layer.