Uncertainty-aware reconstruction of full process maps in laser-directed energy deposition using scarce data

Accurate representation of bead geometry is essential for establishing reliable process-geometry relationships in laser-directed energy deposition (L-DED). In contrast to existing approaches that focus on constructing complete process maps within a single manufacturing process, this work leverages experimentally derived scalar bead descriptors (width, height, and area) from a source process to enable data-efficient reconstruction of process maps for a different target manufacturing process. To achieve this, a neural network-based data-driven modeling approach is proposed, leveraging knowledge from experimental laser wire-DED (LW-DED) to improve predictions for laser powder-DED (LP-DED). The framework is trained using data from 44 LW-DED SS316L and 20 LP-DED IN718 deposits. The approach integrates a pretrained source model with a preprocessor bridging LW-DED and LP-DED process parameters and updates only a subset of parameters to leverage previously learned representations. Monte Carlo dropout is used to quantify predictive uncertainty. Achieving an R 2 of 0.95, coefficient of variation of 5.13%, and 95% coverage of prediction intervals in a fourfold test, the model outperforms baseline approaches with limited adaptation. Reconstructed process maps for LP-DED, derived from the predicted scalar geometry descriptors, show strong physical consistency. Cross-process comparison further reveals distinct deposition behaviors, highlighting the framework’s ability to enable data-efficient learning while providing insight into underlying process-geometry relationships.