Gaussian process regression with physics-guided features for WLAM bead geometry: Repeated validation, algorithmic baselines, and predictive-interval assessment
Abstract
Wire–laser additive manufacturing (WLAM), also referred to as laser wire-feed directed energy deposition, requires stable bead geometry for dimensional control. This study evaluates Gaussian process regression (GPR) for predicting bead width (WB) and height (HB) from laser power, travel speed, and wire feed speed under a small-data setting. Line energy density and the wire-to-travel ratio are added as physics-guided coordinates; they are deterministic transformations rather than independent observations or governing constraints. One experimental run ( n = 24 ) was used for model development with ten repetitions of five-fold cross-validation, and two later runs ( n = 24 ) were reserved for same-system evaluation. Relative to the three original inputs, feature augmentation reduced repeated-CV GPR RMSE by 5.5% for WB and 24.8% for HB. Augmented GPR achieved repeated-CV R 2 = 0.813 ± 0.026 for WB and 0.957 ± 0.007 for HB; after refitting on all development data, later-run R 2 values were 0.742 and 0.959, respectively. Linear regression and a shallow neural network were competitive for several point metrics, showing that the improvement cannot be attributed to GPR alone. Nominal 95% GPR prediction intervals covered 91.7% of WB and 95.8% of HB later-run observations. The contribution is therefore a controlled assessment combining feature ablation, identical-fold algorithmic baselines, repeated resampling, later-run testing, and predictive-interval coverage assessment. The demonstrated scope is limited to measurable stable tracks in one SS316L machine/process window; broader transfer and experimentally confirmed inverse parameter selection remain future work.
Related
- Assessment of DRACCAR predictive capabilities during bundle reflood under low-power conditions: A systematic uncertainty and sensitivity analysis
- Development and validation of a finite element ski helmet model for the biomechanical assessment of oblique head impacts
- Zero-shot time-series foundation models against the Khepri v1.5 baselines for 96-hour carbon-intensity forecasts in nine Nordic bidding zones: a preregistered comparison
- Algorithmic overlaps in the Baxter–Wu model: cluster dynamics under Novotny–Evertz updates
- Algorithmic Randomness and Physical Typicality
BibTeX
@article{bui2026gaussian,
title = {{Gaussian process regression with physics-guided features for WLAM bead geometry: Repeated validation, algorithmic baselines, and predictive-interval assessment}},
author = {Minh-Hien Bui and Dinh-Son Nguyen and Frédéric Vignat},
journal = {Next Materials},
year = {2026},
month = sep,
doi = {10.1016/j.nxmate.2026.103583},
url = {https://hal.science/hal-05766676v1},
note = {CC BY-NC-ND},
}
RIS
TY - JOUR
TI - Gaussian process regression with physics-guided features for WLAM bead geometry: Repeated validation, algorithmic baselines, and predictive-interval assessment
AU - Minh-Hien Bui
AU - Dinh-Son Nguyen
AU - Frédéric Vignat
PY - 2026
DA - 2026/09/28
JO - Next Materials
DO - 10.1016/j.nxmate.2026.103583
UR - https://hal.science/hal-05766676v1
AB - Wire–laser additive manufacturing (WLAM), also referred to as laser wire-feed directed energy deposition, requires stable bead geometry for dimensional control. This study evaluates Gaussian process regression (GPR) for predicting bead width (WB) and height (HB) from laser power, travel speed, and wire feed speed under a small-data setting. Line energy density and the wire-to-travel ratio are added as physics-guided coordinates; they are deterministic transformations rather than independent observations or governing constraints. One experimental run ( n = 24 ) was used for model development with ten repetitions of five-fold cross-validation, and two later runs ( n = 24 ) were reserved for same-system evaluation. Relative to the three original inputs, feature augmentation reduced repeated-CV GPR RMSE by 5.5% for WB and 24.8% for HB. Augmented GPR achieved repeated-CV R 2 = 0.813 ± 0.026 for WB and 0.957 ± 0.007 for HB; after refitting on all development data, later-run R 2 values were 0.742 and 0.959, respectively. Linear regression and a shallow neural network were competitive for several point metrics, showing that the improvement cannot be attributed to GPR alone. Nominal 95% GPR prediction intervals covered 91.7% of WB and 95.8% of HB later-run observations. The contribution is therefore a controlled assessment combining feature ablation, identical-fold algorithmic baselines, repeated resampling, later-run testing, and predictive-interval coverage assessment. The demonstrated scope is limited to measurable stable tracks in one SS316L machine/process window; broader transfer and experimentally confirmed inverse parameter selection remain future work.
SN - 2949-8228
ER -