Directional mesh sensitivity and neural-operator defect correction for weld pool prediction in laser manufacturing
Melt-pool geometry in laser manufacturing is set by three-dimensional heat, flow and solute transport. Finite-volume simulation is the reference tool, but equal refinement in all three directions is prohibitive, and pool quantities are not equally sensitive to each direction. A directional resolution audit is therefore carried out on a finite-volume melt-pool model. The material enters through its thermophysical properties, and the numerical experiments use experimentally validated parameters. On 156 converged cases, doubling the number of $J$ nodes reduces the melt-width error by $1.27\%$ and the liquid-volume error by $2.39\%$; the same refinement of $I$ or $K$ yields only $0.03\%$ and $0.16\%$, about a $40$-fold difference. This matches the Marangoni-dominated free-surface viscous shear layer. A Fourier neural operator then supplies the coarse-to-fine correction. On 48 unseen cases on the $73\times 8\times 20$ mesh, the melt-width error is reduced by $53.9\%$ and the liquid-volume error by $59.6\%$. Resolution should therefore follow directional sensitivity rather than equal refinement. With the conservative solve kept on an affordable mesh and a neural operator restoring the resolution that governs pool geometry, the computational budget is reduced substantially.
Authors
- Xuyang Chen (ORCID: https://orcid.org/0009-0003-2122-6316)
- Jiao Li (ORCID: https://orcid.org/0000-0002-2258-8736)
- Hanwei Wang (ORCID: https://orcid.org/0000-0003-2755-2000)
- Lele Hu
- Wenbo Zhou
- Heng Jiang
Institutions
- National University of Singapore (SG)
- Chinese Academy of Sciences (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-20
- DOI
- https://doi.org/10.21203/rs.3.rs-11359433/v1
- Primary Topic
- Welding Techniques and Residual Stresses
- Type
- preprint