ML Surrogate Model for Binder Jet Distortion Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for quantifying distortion of binder jet parts during sintering are computationally costly and inefficient, as they require extensive simulations and retraining for each geometry, limiting their predictive power and scalability.

Innovation Solution

The development of machine-learning surrogate models, specifically deep neural networks, that can predict distortion and variability of complex geometries by training on simple geometries, using image-based representations and convolutional neural networks (CNNs), enabling rapid prediction and analysis of distortion outputs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If calibrated computer simulation code is used to quantify distortion variability, then measurement precision is improved, but computational cost increases significantly

Engineering Contradiction:
Improvedistortion variability quantificationVSAvoidcomputational cost
Core Design Contradiction:
Measurement precisionVSLoss of energy

Solution Approach 1:

The system performs preliminary actions by training the machine learning model on a dataset of simulated green body parts with known distortion characteristics before actual production. This pre-training phase captures the relationship between manufacturing parameters and distortion outcomes, enabling rapid prediction without requiring costly real-time simulations for each new design.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system creates a digital copy or surrogate model of the complex sintering process through machine learning. Instead of running computationally expensive physics-based simulations for each prediction, the system uses the trained ML model that replicates distortion behavior, providing accurate predictions at a fraction of the computational cost.

Inventive Principle:
Principle #26Copying

2Manufacturing precision

If extensive simulations are performed for each geometry, then manufacturing precision is improved, but productivity decreases

Engineering Contradiction:
Improvedistortion prediction accuracyVSAvoidprediction speed
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The system performs the computationally intensive work of learning distortion patterns in advance through training on diverse green body part geometries and material properties. Once trained, the model provides rapid predictions for new designs without requiring extensive simulations, thus maintaining precision while dramatically improving productivity.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system changes the computational parameters from running full physics-based simulations to using a pre-trained machine learning model. This parameter change transitions the system from high-computation/low-speed mode to low-computation/high-speed mode while preserving prediction accuracy through the model's learned relationships.

Inventive Principle:
Principle #35Parameter changes

3Adaptability or versatility

If retraining is performed for each geometry, then adaptability is improved, but loss of time increases

Engineering Contradiction:
Improvegeometry-specific predictionVSAvoidretraining time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The machine learning model is designed with universality to handle multiple geometry types, material properties, and manufacturing parameters within a single model. The model learns generalizable patterns from diverse training data that apply across different green body part configurations, eliminating the need for geometry-specific retraining while maintaining adaptability.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The system performs comprehensive training in advance on a wide variety of geometries and conditions to build a robust, generalizable model. This preliminary action ensures the model can adapt to new geometries without requiring retraining, as it has already learned the underlying physical relationships that govern distortion across different configurations.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4528572A1System and method for prediction binder jet distortion and variability using machine learning
Publication Date: 2025.03.26 GENERAL ELECTRIC CO
  • EP4528572A1 patent drawingFigure 1
  • EP4528572A1 patent drawingFigure 2
  • EP4528572A1 patent drawingFigure 3

AI summary

A method for predicting distortion of a part during sintering in an additive process includes receiving a computerized representation of a complex geometric part, discretizing the computerized representation of the complex geometric part into a plurality of elements, processing the plurality of elements of the computerized representation of the complex geometric part with a machine-learning model configured to predict a distorted geometry of the complex geometric part in response to a sintering process, wherein the machine-learning model is trained to predict distortion of a set of primitive geometric coupons represented by image data fed into the machine-learning model during training, the set of primitive geometric coupons having fewer geometries than the complex geometric part, the complex geometric part comprises a plurality of geometries corresponding to geometries associated with the set of primitive geometric coupons, and generating a computerized representation of the predicted distorted geometry of the complex geometric part.