Hybrid Multimodal Surrogate Modeling and Uncertainty-Aware Co-Design for L-PBF Ti-6Al-4V with Nanomaterials-Informed...

Area of Science:

• Materials Science
• Additive Manufacturing
• Computational Materials Science

Background:

• Laser powder bed fusion (L-PBF) property prediction is challenged by small, incomplete datasets.
• Missing morphology descriptors limit generalization and co-design in L-PBF.

Purpose of the Study:

• To develop a hybrid multimodal surrogate strategy for L-PBF property prediction with missing data.
• To enable constraint-aware candidate generation and process selection.

Main Methods:

• Coupling engineered process physics features with morphology proxies using a two-stage embedding module.
• Employing gradient-boosted tree regressors with missingness-aware feature filtering and imputation.
• Utilizing 5-fold GroupKFold cross-validation grouped by set_id.

Main Results:

• Achieved high accuracy (RMSE/R²: 11.07 MPa/0.895 for yield, 13.88 MPa/0.873 for UTS, 0.677%/0.861 for elongation, 2.38 GPa/0.663 for modulus).
• Surface roughness and hardness prediction remain challenging (R² ~0.12-0.11).
• Enabled identification of manufacturing recipes balancing strength and surface objectives under uncertainty.

Conclusions:

• The hybrid surrogate strategy offers a practical blueprint for small-data additive manufacturing studies.
• The framework can be extended for richer microstructure data and prospective validation.
• Accelerates functional and biomedical alloy development through improved prediction and selection.