Machine learning-based microstructure prediction during laser sintering of alumina.

The primary aim of this research is to develop a machine learning model capable of predicting microstructure evolution during laser sintering of alumina. The study seeks to overcome the limitations of traditional methods that require detailed knowledge of governing physical laws. The researchers propose a new algorithm called RCWGAN-GP, which combines regression and generative adversarial networks. This approach allows for prediction without explicit knowledge of the underlying physics. The study focuses on laser sintering of alumina, a widely used ceramic material. The goal is to generate realistic microstructure images under both trained and unexplored conditions. The researchers also aim to validate the algorithm's accuracy using both experimental and simulated datasets. By achieving this, the study contributes to the development of data-driven predictive tools in materials science.

The RCWGAN-GP algorithm was developed using a combination of regression and conditional generative adversarial networks. This approach incorporates a Wasserstein loss function and gradient penalty to improve training stability. The algorithm was trained on experimental scanning electron microscopy (SEM) images of laser-sintered alumina. These images were collected under various laser power settings to capture microstructural variations. In addition to experimental data, the researchers used computer-generated datasets based on the Johnson-Mehl-Avrami (JMA) equation. These simulated datasets represent secondary-phase growth during sintering. The model was tested for its ability to regenerate micrographs at trained laser powers and predict structures under unexplored conditions. Quantitative metrics were used to assess the accuracy of predicted microstructure features, including grain morphology and spatial distribution. The algorithm's performance was evaluated against both experimental and simulated datasets.

The RCWGAN-GP algorithm successfully regenerated SEM micrographs of laser-sintered alumina under trained laser powers. The generated images closely matched the experimental data in terms of particle morphology and pore distribution. The algorithm also accurately predicted microstructure under unexplored laser power conditions. These predictions showed strong agreement with new experimental SEM images. The study tested the algorithm using simulated datasets based on the JMA equation. The model accurately regenerated micrographs at trained time points, capturing grain shapes, sizes, and spatial distributions. The predicted secondary phase fraction followed the JMA curve with high accuracy. Quantitative analysis confirmed that the model's predictions aligned closely with both experimental and simulated data. These results suggest that the RCWGAN-GP algorithm is effective for microstructure prediction in laser sintering processes.

The RCWGAN-GP algorithm demonstrates strong potential for predicting microstructure during laser sintering of alumina. The model accurately regenerates SEM micrographs under trained conditions and reliably predicts structures under unexplored laser powers. The algorithm's performance was validated using both experimental and simulated datasets. The predicted microstructure features, including particle morphology and pore distribution, closely match experimental observations. The model also accurately follows the JMA curve for secondary-phase growth. These findings suggest that the RCWGAN-GP algorithm is a valuable tool for microstructure prediction in materials science. The study highlights the effectiveness of machine learning approaches in capturing complex microstructural evolution. The results support the use of data-driven methods for predictive modeling in advanced manufacturing processes.