Fast and accurate prediction of temperature evolutions in additive manufacturing process using deep learning

Abstract

Typical computer-based parameter optimization and uncertainty quantification of the additive manufacturing process usually requires significant computational cost for performing high-fidelity heat transfer finite element (FE) models with different process settings. This work develops a simple surrogate model using a feedforward neural network (FFNN) for a fast and accurate prediction of the temperature evolutions and the melting pool sizes in a metal bulk sample (3D horizontal layers) manufactured by the DED process. Our surrogate model is trained using high-fidelity data obtained from the FE model, which was validated by experiments. The temperature evolutions and the melting pool sizes predicted by the FFNN model exhibit accuracy of 99% and 98%, respectively, compared with the FE model for unseen process settings in the studied range. Moreover, to evaluate the importance of the input features and explain the achieved accuracy of the FFNN model, a sensitivity analysis (SA) is carried out using the SHapley Additive exPlanation (SHAP) method. The SA shows that the most critical enriched features impacting the predictive capability of the FFNN model are the vertical distance from the laser head position to the material point and the laser head position.

Impact of Boundary Parameters Accuracy on Modeling of Directed Energy Deposition Thermal Field

“…Some future work will be done about the influence of the thickness of the deposit on the temperature field for 3D cases compared to their 2D model. These interactions will be analyzed, thanks to artificial intelligence, as it has already been done previously in [51] to reproduce the different results with a surrogate model that can be rapidly trained thanks to the 2D assumption.…”

An explainable machine learning-based probabilistic framework for the design of scaffolds in bone tissue engineering

“…In this work once the ML-ROM is evaluated, the Shapley additive explanations (SHAP) method [62] is employed for sensitivity analysis, using the essential dataset extracted through the ROM. The SHAP value analysis proposed recently in the frame of explainable AI techniques (XAI) [63] and applied in a variety of applications in the fields of material engineering and computational mechanics [64][65][66][67], is used in this work to increase the transparency and explainability of the ML-ROM predictions by evaluating the contribution of each parameter on the cell differentiation stimulus.…”

Optimizing laser power of directed energy deposition process for homogeneous AISI M4 steel microstructure

“…Zhang et al [47] applied two machine learning algorithms to predict melting pool temperature in a DED process with high accuracy. However as pointed by Pham et al [48,49] an accurate deep learning strategy requires significant experimental data or validated finite element simulation results of the process to generate accurate predictions. In Pham et al [48], a simple FFNN architecture was applied to model DED manufacturing of AISI M4 samples.…”

“…However as pointed by Pham et al [48,49] an accurate deep learning strategy requires significant experimental data or validated finite element simulation results of the process to generate accurate predictions. In Pham et al [48], a simple FFNN architecture was applied to model DED manufacturing of AISI M4 samples. The developed surrogate model was 180 times quicker than a FE simulation and it allowed to perform an uncertainty quantification and sensitivity analysis of M4 DED sample manufacturing process.…”