Machine learning-based prediction of directed energy deposition process with small-size experimental data
The inconstancy in material properties and complex geometry make directed energy deposition (DED) a difficult process to control and optimize. Understanding the process-structure-property relationship and modeling geometry from track, to layer and multi-layer are keys to advancing DED. Experimentati...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Thesis-Doctor of Philosophy |
Language: | English |
Published: |
Nanyang Technological University
2024
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/173602 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | The inconstancy in material properties and complex geometry make directed energy deposition (DED) a difficult process to control and optimize. Understanding the process-structure-property relationship and modeling geometry from track, to layer and multi-layer are keys to advancing DED. Experimentation and modeling face challenges like costs, computational demands and inaccuracies in complex geometry. This study introduces machine learning solutions to these challenges through four main contributions:
Contribution One: Design of Experiments Informed Deep Learning for Modeling of Directed Energy Deposition with Small-sized Experimental Dataset
A design of experiments (DOE) informed deep learning (DL) model is developed for modeling of laser powder-based DED process to address the challenge of high throughput data generation in the DED process development. A small-size experimental dataset is obtained according to DOE, by which a large-size dataset is augmented via the DOE regression model and then used to pre-train the DL model. A subset of experimental data is employed to fine-tune the DL model. The presently developed DOE-informed DL model is validated via single-track deposition of stainless steel 316L, in which the cross-section geometrical characteristics (including width, height, dilution depth), are predicted with R2 value of 0.73, 0.92, and 0.81. Meanwhile, the predicted ranges for the porosity and hardness become the subset of the experimental results, namely [0.05%, 0.25%] belongs to [0.04%, 0.30%] for porosity, and [170 HV, 180 HV] belongs to [168 HV, 182 HV] for hardness.
Contribution Two: Contrastive Learning Based on Hierarchical Graph of Microstructures through Directed Energy Deposition Process to Establish Process-Structure-Property Relationship via Autoencoder
An innovative algorithm is developed to transform microstructures to hierarchical graphs without feature engineering. The hierarchical graph comprises two layers. Initially, pixel data, which includes Euler angles, phase, and position, forms the pixel-wise graphs within individual grains. Following this foundation, the grains serve as nodes, constituting the second layer. Importantly, the hierarchical graph preserves essential measurement data and structural details for training machine learning models. After that, a contrastive learning model based on hierarchical graph is designed to capture representations of microstructures obtained by the electron backscatter diffraction (EBSD) technique. Using the learned microstructure representations, an autoencoder model for DED is developed to establish the relationship among process parameters, microstructures, and material properties, completing the process-structure-property cycle quantitatively. The mean absolute percentage error for yield strength, ultimate tensile strength, and hardness predictions are all under 5.2%. The prediction of porosity follows the trend with small root mean square error (0.3%). While, the predictions of elongation (30% to 45%) and Young’s modules (120 to 140 GPa) become smooth without capturing the spike values.
Contribution Three: Using Track-wise Long Short-Term Memory for Multi-layer Geometry Predicting in Directed Energy Deposition Process
A track-wise long short-term memory (LSTM) machine learning algorithm is proposed for multi-layer geometry forecasting in DED process. Process parameters can be assigned to each track-wise deposition and LSTM takes account of the effects of previous tracks. The Y-value of experimental points are predicted with R2 value of 0.8 and the height is predicted with R2 value of 0.85. In the future work, hatching spacing can be optimized with minimum waviness by layer-wise and Z-increment value can be optimized with nearly desirable height by multi-layer machine learning models separately.
Contribution Four: Utilizing Simulation Data for Pre-training and Experimental Data for Fine-tuning in the Prediction of the DED Process with Uncontrollable Factors
Integration of simulation data generated from physics-based numerical simulation for pre-training and experimental data for fine-tuning machine learning framework is proposed to further reduce the number of experiments and incorporate the in-process together with pre-process and post-process information. A medium-sized dataset is generated by a physics-based model according to DOE, by which a large-size dataset is augmented via the DOE regression model and then used to pre-train the DL model. A subset of simulation data is employed to fine-tune the DL model to let the DL model mimic the simulation behavior. Finally, the transfer learning technique is applied to the fine-tuned DL model with experimental data. The presently developed simulation data for pre-training and experimental data for fine-tuning model is validated via single-track deposition of stainless steel 316L with uncontrollable factor sulfur content, in which the cross-section dilution Y-value and the bead Y-value are predicted with R2 value 0.49 and 0.95. The simulation data for pre-training and experimental data for fine-tuning model developed is based on a single-track deposition; in future work, it will be extended to multi-layer deposition. |
---|