Researchers have developed a mechanism-data fusion framework integrating finite element modeling, phase-field modeling, and dual-level long short-term memory networks for rapid process-microstructure prediction and optimization in laser directed energy deposition of Ti-6Al-4V. Published in Advanced Equipment, this framework enables efficient prediction of thermal and microstructural evolution and provides a promising approach for microstructure-process optimization in metal additive manufacturing.
Laser directed energy deposition (L-DED) offers significant potential for manufacturing. However, repeated thermal cycles and rapid solidification during layer-by-layer deposition lead to complex thermal histories and microstructural evolution, making it difficult to efficiently establish relationships between processing parameters and microstructures. Addressing this challenge, researchers from Yanshan University, in collaboration with researchers from Prairie View A&M University and the University of Michigan-Dearborn, have developed a mechanism-data fusion framework integrating multiscale physical modeling with a dual-level long short-term memory (LSTM) surrogate model for rapid prediction and optimization of microstructure evolution in L-DED Ti-6Al-4V.
The framework combines the finite element method (FEM) and phase-field method (PFM) to establish a multiscale physical model. The FEM predicts temperature fields during deposition, while the PFM uses the calculated thermal histories to simulate β-grain nucleation, competitive growth, and epitaxial evolution. The resulting simulation data are then used to train a dual-level LSTM surrogate model, linking processing parameters to thermal responses and ultimately to microstructural characteristics.
The multiscale model was validated against experimental measurements. The predicted melt-pool morphology and thermal histories showed good agreement with experimental observations. EBSD characterization further confirmed the formation of epitaxially grown columnar β grains, with simulated grain dimensions consistent with experimentally measured ranges. These comparisons support the ability of the coupled FEM-PFM model to reproduce the major characteristics of thermal and microstructural evolution during L-DED.
The simulations revealed clear relationships between processing parameters and β-grain evolution. Increasing scanning speed generally reduced the average β-grain area and increased the number of grains, whereas increasing laser power promoted grain coarsening. These trends were associated with changes in thermal history, which influence grain nucleation and competitive epitaxial growth.
To capture the temporal dependence introduced by layer-wise thermal accumulation, the dual-level LSTM architecture uses a first-level surrogate model to predict thermal evolution from processing parameters and a second-level model to predict microstructural evolution from thermal information. Cross-validation yielded R² of 0.96 and 0.95 for the thermal and microstructure surrogate models, respectively. Under four testing conditions not included in the training dataset, the framework maintained an average R² of approximately 0.95 and an average relative error of about 8.14%, while outperforming a conventional multilayer perceptron model in layer-wise microstructure prediction.
The framework also reduced computational cost. A complete FEM-PFM simulation for one processing condition required more than 40 hours, whereas the surrogate framework accelerated new process-parameter evaluations by approximately 2.4 × 10² times. It was further used to rapidly explore processing conditions associated with refined and near-equiaxed β-grain morphologies, demonstrating its potential for efficient process optimization.
Experimental measurements under six processing conditions further supported the framework, yielding an R² of 0.96 and an average relative error of 11.84%. Additional phase-field simulations outside the training domain showed that the model could capture the overall trend of grain evolution, although its prediction accuracy decreased. While further validation across broader processing windows is required, the framework provides a promising approach for integrating physical modeling and machine learning to accelerate microstructure prediction and process optimization in metal additive manufacturing.
This paper “A mechanism-data fusion framework for process-microstructure prediction and optimization in laser directed energy deposition of Ti-6Al-4V using dual-level surrogate modeling” was published in Advanced Equipment .
Zhang Y, Zhao T, Liu P, Zhou X, Wang X, Yang J, et al . A mechanism-data fusion framework for process-microstructure prediction and optimization in laser directed energy deposition of Ti-6Al-4V using dual-level surrogate modeling. Advanced Equipment . 2026(2):0006, https://doi.org/10.55092/ae20260006.
Advanced Equipment
Computational simulation/modeling
Not applicable
A mechanism-data fusion framework for process-microstructure prediction and optimization in laser directed energy deposition of Ti-6Al-4V using dual-level surrogate modeling
25-Aug-2026