What question did this study set out to answer?

The aim is to establish a relationship between LPBF process parameters and resulting microstructures using deep learning.

April 1, 2026Open Access

Process-structure linkage for laser powder bed fusion using simulated microstructure images and deep learning

Key Points

The aim is to establish a relationship between LPBF process parameters and resulting microstructures using deep learning.
Developed a data-driven modelling framework using simulated microstructure images.
Used laser energy density as the primary input parameter.
Extracted features like grain boundaries and texture for classification.
Employed a convolutional neural network optimized with hyperparameter tuning.
Integrated transfer learning to enhance model training and efficiency.
Achieved a training accuracy of approximately 94%.
Obtained a microF1 score of 0.92 for microstructure classification.
Demonstrated strong generalization with real LPBF-processed samples through validation.

Abstract

Abstract Understanding the influence of process parameters on resulting microstructures is crucial for establishing process-structure linkage in metal additive manufacturing. Recently, the data-driven modelling approach has been incorporated in various advanced processes such as laser powder bed fusion (LPBF) due to its inherent ability to solve complex multiphysics problems. This study presents a data-driven modelling framework assisted by simulated microstructure data to classify the microstructures based on input LPBF parameters. Laser energy density is employed as the primary process variable, while grain boundaries, edges, texture, and average grain size are extracted as features to train a deep learning model. A thermal–grain growth numerical model, developed using a coupled finite element and cellular automata framework, is used to simulate the evolution of microstructures under varying laser energy densities. Data augmentation is used to alleviate big data dependency, and an LPBF microstructure dataset of 405 images is developed, consisting of three classes. A novel convolutional neural network (CNN) is developed and trained from scratch, optimized through extensive hyperparameter tuning. To further improve training efficiency and model generalization, transfer learning is integrated using the VGG16 architecture. The resulting model achieved a high training accuracy of ~ 94% and a microF1 score of 0.92, demonstrating the effectiveness of transfer learning in accelerating convergence and improving performance in microstructure classification. Model interpretability is further investigated using t-distributed stochastic neighbour embedding and Shapley Additive Explanations, revealing that the CNN learns class-distinctive features aligned with known microstructural characteristics. The trained model is further validated using experimentally acquired micrographs, demonstrating strong generalization and reliable classification performance across real LPBF-processed samples.

Perguntar à IA

Bookmark

View Full Paper