Prediction of weld pool and keyhole geometries in high-power laser beam welding through a physics-informed generative artificial intelligence approach

  • The weld pool and keyhole geometries are critical characteristics in evaluating the stability of the high-power laser beam welding (LBW) process and determining the resultant weld quality. However, obtaining these data through experimental or numerical methods remains challenging due to the difficulties in experimental measurements and the high computational demands of numerical modelling. This paper presents a physics-informed generative approach for predicting weld pool and keyhole geometries in the LBW process. With the help of a well experimentally validated numerical model considering the underlying physics in the LBW, the geometries of the weld pool and keyhole under various welding conditions are calculated, serving as the dataset of the generative model. A Conditional Variational Autoencoder (CVAE) model is employed to generate realistic 2D weld pool and keyhole geometries from the welding parameters. We utilize a β-VAE model with the Evidence Lower Bound (ELBO) loss functionThe weld pool and keyhole geometries are critical characteristics in evaluating the stability of the high-power laser beam welding (LBW) process and determining the resultant weld quality. However, obtaining these data through experimental or numerical methods remains challenging due to the difficulties in experimental measurements and the high computational demands of numerical modelling. This paper presents a physics-informed generative approach for predicting weld pool and keyhole geometries in the LBW process. With the help of a well experimentally validated numerical model considering the underlying physics in the LBW, the geometries of the weld pool and keyhole under various welding conditions are calculated, serving as the dataset of the generative model. A Conditional Variational Autoencoder (CVAE) model is employed to generate realistic 2D weld pool and keyhole geometries from the welding parameters. We utilize a β-VAE model with the Evidence Lower Bound (ELBO) loss function and include Kullback-Leibler divergence annealing to better optimize model performance and stability during training. The generated results show a good agreement with the ground truth from the numerical simulation. The proposed approach exhibits the potential of physics-informed generative models for a rapid and accurate prediction of the weld pool geometries across a diverse range of process parameters, offering a computationally efficient alternative to full numerical simulations for process optimization and control in laser beam welding processes.zeige mehrzeige weniger

Volltext Dateien herunterladen

  • Manuscript(10.06.2025).pdf
    eng

Metadaten exportieren

Weitere Dienste

Suche bei Google Scholar Anzahl der Zugriffe auf dieses Dokument
Metadaten
Autor*innen:Xiangmeng Meng, Marcel Bachmann, Pascal Kising, Fan Yang, Michael RethmeierORCiD
Dokumenttyp:Beitrag zu einem Tagungsband
Veröffentlichungsform:Graue Literatur
Sprache:Englisch
Titel des übergeordneten Werkes (Englisch):Proceedings of the ICALEO 2025, 44th International Congress on Applications of Lasers & Electro-Optics
Jahr der Erstveröffentlichung:2025
Organisationseinheit der BAM:9 Komponentensicherheit
9 Komponentensicherheit / 9.3 Schweißtechnische Fertigungsverfahren
Erste Seite:1
Letzte Seite:10
DDC-Klassifikation:Technik, Medizin, angewandte Wissenschaften / Ingenieurwissenschaften / Ingenieurwissenschaften und zugeordnete Tätigkeiten
Freie Schlagwörter:Generative artificial intelligence; Laser beam welding; Machine learning; numerical simulation
Themenfelder/Aktivitätsfelder der BAM:Material
Material / Degradationsmechanismen
Veranstaltung:International Congress of Applications of Lasers & Electro-Optics 2025
Veranstaltungsort:Orlando, USA
Beginndatum der Veranstaltung:12.10.2025
Enddatum der Veranstaltung:16.10.2025
Verfügbarkeit des Dokuments:Datei im Netzwerk der BAM verfügbar ("Closed Access")
Datum der Freischaltung:10.12.2025
Referierte Publikation:Nein
Einverstanden
Diese Webseite verwendet technisch erforderliche Session-Cookies. Durch die weitere Nutzung der Webseite stimmen Sie diesem zu. Unsere Datenschutzerklärung finden Sie hier.