Directed Energy Deposition (DED) is widely used for fabricating metallic components, where thermal history strongly influences microstructure, residual stress, and defect formation. Accurate thermal modeling is therefore very important for process optimization and in-situ process control. Traditional analytical models offer computational efficiency but rely on simplifying assumptions that limit accuracy, while finite element simulations provide higher fidelity at the expense of a high computational cost. Conventional machine learning (ML) models, though capable of capturing nonlinear behav ior, require large datasets that are costly to obtain in additive manufacturing, which limits their application in real-world industries. To overcome these limitations, this study introduces a hybrid Physics-Informed Machine Learning (PIML) framework that integrates an analytical thermal model with ML. Our methodology first employs a computationally effi cient analytical model, based on a transient moving point heat source solution, to generate a rapid, physics-based initial estimate of the temperature field. This analytically estimated temperature is then integrated as a direct input feature into the ML model, alongside process parameters and material properties, enabling the learning of residual thermal effects neglected by the analytical formulation. Validation against experimental data collected from the literature shows that while the pure analytical model achieves a coefficient of determination (R²) of 0.73 with a root mean square error (RMSE) of 244 °C, the suggested PIML approach significantly improves accuracy. Among the tested algorithms (Support Vector Regression, Least Squares Boosting, and Random Forest), the latter achieved the best performance, with R² = 0.98 and RMSE = 20.7 °C. Notably, this accuracy was obtained using a limited number of experimental data points, highlighting the potential of physics-based ML to reduce data requirements while maintaining generalization across multiple alloys (Ti-6Al-4 V, stainless steel, AlSi10Mg, etc.), even for unseen test datasets. The findings underline the dual advantages of the proposed framework: enhanced predictive capability and a considerable reduction in data dependency. This hybrid modeling approach offers a promising pathway toward efficient digital twins for DED, enabling faster process optimization and more reliable implementation in metallic additive manufacturing
