Towards real-time prediction of thermal history and hardness in laser powder bed fusion using deep learning

Laser powder bed fusion (PBF-LB) of quenched-and-tempered steels is governed by highly localized thermal histories that control microstructure evolution and hardness. Predicting these process–structure–property relationships typically requires computationally intensive finite element (FE) simulations, limiting real-time applicability. Here, we present a physics-informed deep learning surrogate for real-time prediction of thermal histories and hardness in PBF-LB of AISI 4140. The model integrates process parameters and spatial laser–material interactions within an autoregressive sequence framework to capture path-dependent thermal behavior. Trained on multiscale FE data, the model reconstructs local temperature–time histories with high fidelity and enables hardness prediction via a non-isothermal Hollomon–Jaffe relationship. A two-layer LSTM ensemble achieves a temperature RMSE of (2.6 ± 1.2) K on an independent test set. The central contribution is the prediction of complete local thermal histories and their propagation through a validated tempering model to obtain local hardness under varying process conditions and cross-section geometries. ... mehrBeyond this specific system, the proposed framework establishes a generalizable strategy for bridging high-fidelity simulation and real-time prediction in manufacturing processes. It thus provides a foundation for accelerated process design, in situ decision-making, and the realization of digital twins in materials processing.