Skip to Content

MeltPoolGAN: Auxiliary Classifier Generative Adversarial Network for melt pool classification and generation of laser power, scan speed and scan direction in Laser Powder Bed Fusion

The study presents MeltPoolGAN, a machine learning architecture designed for the classification and controlled generation of melt pool images in additive manufacturing. It significantly improves the accuracy of classifying process parameters such as laser power, scan speed and scan direction, achieving around 97% accuracy for power and scan speed classification and less than 3 degrees of error for scan direction estimation.


Type of Work: Scientific Publication

Main Author: Jan Petrik

Affiliation: ETH Zurich

Co-Authors: Baris Kavas, Markus Bambach

Date: 10th November 2023

Journal: Journal of Additive Manufacturing

Online: ScienceDirect

MeltPoolGAN employs an Auxiliary Classifier Generative Adversarial Network (ACGAN) framework, incorporating a novel approach to processing and analyzing melt pool images within the additive manufacturing domain. The methodology includes the use of generative adversarial networks (GANs) enhanced with auxiliary classifiers to accurately identify and classify process parameters such as laser power, scan speed, and scan direction across an extensive range of 371 classes. This setup allows for high classification accuracies and the generation of realistic melt pool images under varying conditions. The research used two different datasets, NIST and an in-house generated ETH dataset, to test the robustness and versatility of the model.
The MeltPoolGAN architecture demonstrates remarkable performance, with classification accuracies of around 97% for power and scan speed classes, and scan direction estimation errors of less than 3 degrees. Such results in terms of rotation have been achieved thanks to the loss function developed, specifically the Gaussian Cross Entropy loss function. Furthermore, the ability to generate controlled and accurate melt pool images opens up new avenues for data generation, classifier training and offline process parameter optimisation. Ultimately, it is believed that this AI architecture can be used in the future for real-time control of the LPBF process to check for anomalies and improve the overall geometric quality of the printed part.


  • Image 1
    The Aconity Midi+ system was used to create the ETH dataset by printing on a build platform and capturing melt pool images using a CMOS camera.
  • Image 2
    MeltPoolGAN's workflow includes a generator that creates images from noise and laser parameters (power, direction, speed) and a discriminator that evaluates if melt pool images are real or fake, identifying their laser settings. If the printing process aligns with expectations, predicted and actual parameters will match.
  • Image 3

    Left: MeltPoolGAN generates melt pool images across 25 categories with a single rotation, covering a power range of 110 to 190 W and scan speeds from 200 to 1000 mm/s.

    Right: Predictions of melt pools along the laser path, with scan speeds v1 at 900 mm/s and v2 at 600 mm/s.