Work

Grid-Optimized Neural Operators (GONOs) for FE-Node-Consistent Thermo-Mechanical Field Reconstruction in Electron-Beam Powder-Bed Fusion.

Year: 2026
Type: article
Abstract: High-fidelity finite-element (FE) simulations of electron-beam powder-bed fusion (EB-PBF) provide thermo-mechanical fields but are too expensive for repeated field queries. This work proposes grid-optimized neural operators (GONOs), which couple FE nodal data with structured-grid neural operators through a bidirectional point-cloud–grid transformation. The interface projects irregular FE-node fields onto an auxiliary Cartesian grid, applies the neural operator in grid space, and maps the predictions back to the original FE nodes, where training and evaluation are performed. \changeblue{This retains FE-node-level supervision and evaluation while enabling the use of established structured-grid neural-operator architectures.} The framework is evaluated on a fixed-process, 20-layer Ti-6Al-4V EB-PBF web-spider benchmark. \changeblue{This study addresses one-shot synchronized-state field reconstruction, defined here as ordinal frame-index-aligned cross-layer reconstruction: the Layer~1 temperature field at intra-step frame $s$ is mapped to temperature, von Mises stress and displacement-magnitude fields for Layers~2--20 at index $s$ of their respective sequential FE analysis steps. Because these steps occur at different global build times, the outputs are neither a simultaneous whole-build snapshot nor an autoregressive time forecast.} Across three independent training seeds at the native $41\times41$ grid, the improved FNO, enhanced UNet, and MWT variants achieve mean overall relative $L_2$ errors of 0.01849, 0.01659, and 0.01270, respectively. Tests on irregular quadrilateral and triangular meshes demonstrate that the point-cloud–grid interface remains effective on the tested two-dimensional non-Cartesian FE-node layouts. \changeblue{A resolution study indicates that an auxiliary-grid multiplier of $r=2.5$ provides a favorable accuracy--cost trade-off for this benchmark.} \changeblue{At the native grid, with cached P2G inputs, GONOs reconstruct all 19 layer-wise states for the complete 300-sample test set in approximately 7--19~s on a single GPU, including G2P back-projection.} For comparison, generation of the high-fidelity FE reference simulation required approximately 11 days on 120 CPU cores. These results demonstrate an efficient and FE-node-consistent surrogate framework for rapid thermo-mechanical field reconstruction in EB-PBF.
Language: English
Open Access status: bronze
OpenAlex ID: W7225870871
FWCI:
Cites:
Cited by:
Related to: 10