Data-Driven Scanning Pattern Selection for Multi-Layer Deposition in Laser Directed Energy Deposition via Deep Regression Method with Simulation Integration

Technical Summary: Data-Driven Scanning Pattern Selection for Multi-Layer Deposition in Laser Directed Energy Deposition via Deep Regression Method with Simulation Integration

Problem Statement
In Laser Directed Energy Deposition (LDED), heat accumulation during the multi-layer deposition process significantly affects the thermal field, leading to undesirable microstructural characteristics such as coarse grains, elemental segregation, and brittle phase formation. While process parameters like laser power and speed influence heat conduction, the scanning strategy is a critical factor often constrained by equipment limitations when adjusted directly. Continuous Laser Deposition (CLD) involves continuous energy input, which reduces cooling rates, slows solidification, and promotes the growth of large dendrites and the segregation of alloying elements. Existing methods for optimizing deposition patterns, such as experimental trial-and-error or heuristic approaches, are inefficient and struggle to predict thermal fields for complex, multi-layer scenarios rapidly.

Methodology
This study proposes a data-driven approach to optimize scanning patterns by integrating the Quasi-Continuous Wave (QCW) process with a deep regression algorithm, termed the Interval Pulsed Laser Deposition (IPLD) process. The methodology consists of three primary stages:

1. Finite Element Simulation and Dataset Construction: A multi-layer finite element model was established to simulate the thermal field of a 20-layer thin-walled structure. To ensure generalization, the simulation model utilized random position patterns consisting of three initial single-track deposited layers with 33 deposition positions per layer, generating a total of 1,000 simulation results. The model assumed a thin-walled structure where shape changes during deposition did not significantly alter thermal variations, focusing instead on the thermal influence of adjacent layers.
2. Deep Regression Model Development: A deep regression model based on a Skip-Connected 3D Convolutional Autoencoder (SC-3DCAE) was developed. This architecture, inspired by U-net, utilizes skip connections and 3D convolutions to extract features of heat distribution variations caused by different deposition patterns. The model was trained on the simulated thermal field dataset (3:1 training-to-test split) to predict thermal fields for arbitrary patterns. The model achieved a normalized mean squared error (nMSE) below 0.01 and an accuracy of 97.55% after 3,000 epochs.
3. Pattern Optimization and Validation: The trained model was employed to rapidly screen random deposition patterns to identify those maximizing the cooling rate. A step-by-step selection method was used: the optimal pattern for the first three layers was identified, and the last two layers served as the initial input for the next iteration, continuing until all 20 layers were defined. A "brick-laying" overlapping strategy with a 50% staggered ratio was implemented to prevent gaps and ensure layer fusion. The optimized IPLD process was experimentally validated by depositing DD5 alloy onto an Inconel 718 substrate, with results compared against conventional CLD.

Key Contributions

• Integration of QCW and Deep Learning: The study introduces a novel scanning pattern selection method that combines the discrete deposition zones of the QCW process with a deep regression algorithm (SC-3DCAE) to overcome the limitations of continuous energy input.
• Rapid Thermal Field Prediction: The developed SC-3DCAE model enables millisecond-level prediction of thermal fields for arbitrary multi-layer deposition patterns, bypassing the need for computationally expensive finite element simulations for every candidate pattern.
• Systematic Microstructural Analysis: The research provides a comprehensive comparison of the IPLD and CLD processes, analyzing grain orientation, dendrite spacing, precipitate size, and elemental segregation through Electron Backscatter Diffraction (EBSD), Scanning Electron Microscopy (SEM), and Electron Probe Microanalyzer (EPMA) techniques.

Results
Experimental and simulation results demonstrate that the IPLD process significantly outperforms the CLD process in controlling heat accumulation and microstructure:

• Thermal and Microstructural Characteristics: The cooling rate of the CLD process decreases from approximately 3,500 K/s to 1,500 K/s in the bottom region and drops to 250 K/s in the stable middle region. In contrast, the IPLD process maintains significantly higher cooling rates, ranging from 3,500 K/s to 7,000 K/s in the bottom region and 1,100 K/s to 2,000 K/s in the middle region. Consequently, IPLD samples exhibited finer columnar grains with a stronger <001> orientation and smaller grain widths compared to CLD samples, where heat accumulation led to significant grain coarsening.
• Dendrite and Precipitate Refinement: The IPLD process resulted in smaller primary dendrite spacing and less developed secondary dendrite arms. The proportion of large-sized carbides (>0.8 μm) in IPLD samples was only 2%, compared to 11.2% in CLD samples.
• Mechanical Properties: Tensile testing revealed that while ultimate tensile strengths were comparable (896 MPa for IPLD vs. 891 MPa for CLD), the elongation of the IPLD sample was significantly higher at 34.7% compared to 29.9% for CLD. Fracture analysis showed that IPLD samples possessed smaller fracture dimples and fewer pores, indicating a more uniform stress distribution and better ductility.
• Elemental Segregation: Simulation using the Scheil solidification model confirmed that the higher solidification rate of IPLD reduced the segregation of elements such as Nb, Ti, and Ta in the interdendritic regions, leading to smaller and more numerous carbide precipitates rather than large, detrimental ones.

Significance
The paper claims that the proposed IPLD process, driven by data-scanning pattern optimization, offers a feasible research direction for improving the deposition quality of large-sized parts in LDED. By regulating the scanning pattern rather than relying solely on process parameters, the method achieves high cooling rates and low heat accumulation, effectively controlling grain size, orientation, and precipitate distribution. The study suggests that this approach is particularly valuable for depositing single-crystal or polycrystalline superalloys where thermal management is critical to preventing defects and ensuring mechanical integrity. The integration of deep regression with simulation provides a scalable framework for rapid pattern optimization, addressing the computational bottlenecks associated with traditional multi-layer thermal modeling.