Physics-guided neural network framework for surface roughness prediction in additively manufactured metallic alloys

Surface roughness (Ra) significantly influences fatigue performance, corrosion behavior, and postprocessing requirements in additively manufactured metallic components. In practice, Ra optimization remains dependent on alloy-specific experimental campaigns, limiting scalability and slowing process qualification. To address this challenge, this study develops a physics-guided neural modeling framework for trend-level Ra prediction using physically motivated energy descriptors and data-efficient learning strategies. The approach integrates heat input, energy density, and normalized energy as physics-derived features, together with surrogate-assisted training designed to interpolate experimentally and literature-supported behavior within validated process windows. Separate neural networks were trained for 316L stainless steel, Ti-6Al-4V, and AlSi10Mg using an identical architecture, feature set, and hybrid data–physics loss formulation. The contribution lies in the engineering integration of physics-guided feature construction, controlled surrogate interpolation, and alloy-specific model training within a unified methodological workflow, rather than in proposing a new neural architecture or enforcing governing equations. Despite limited experimental datasets (75 samples for 316L and approximately 20 literature-derived samples per alloy), the framework demonstrates strong cross-validated trend consistency (R2 ≈ 0.98–0.995) under fivefold evaluation. Residual analysis and response-surface visualizations confirm stable and physically plausible behavior. The reported R2 values reflect agreement in dominant process–roughness trends rather than deterministic predictive certainty. Overall, the results indicate that a consistent physics-guided modeling strategy can support interpretable and data-efficient surface roughness prediction within experimentally qualified domains. The framework provides a foundation for extending physics-guided learning to related additive manufacturing quality metrics while remaining confined to validated operating windows.