LPBF Processing Map Predictions of Novel Aluminum Alloys using Hybrid FEM-Multitask Neural Network Approach: FEM dataset and ML ANN models

LPBF Processing Map Predictions of Novel Aluminum Alloys using Hybrid FEM-Multitask Neural Network Approach: FEM dataset and ML ANN models F. Bahari-Sambran1,2, F. Carreño*1, A. Orozco-Caballero1, C.M. Cepeda-Jiménez1 1Physical Metallurgy Department, CENIM-CSIC, Av. Gregorio del Amo 8, 28040 Madrid, Spain 2Department of Mechanical Engineering, Chemistry and Industrial Design, Polytechnic University of Madrid, Ronda de Valencia 3, 28012, Madrid, Spain *Corresponding author This work is part of an investigation regarding Hybrid FEM-Multitask Neural Network Approach for Processing Map Prediction in LPBF of Novel AlFeCrX Alloys F. Bahari-Sambran, F. Carreño*, A. Orozco-Caballero, C.M. Cepeda-Jiménez, performed at Physical Metallurgy Department, CENIM-CSIC, Madrid, Spain, which will be published shortly. The development of predictive models for additive manufacturing (AM) is challenging due to limited experimental data and high costs, especially for novel alloys. This study presents a hybrid modeling approach combining finite element modeling (FEM) and artificial neural networks (ANNs) through transfer learning (TL) to predict melt pool size, defect types (lack-of-fusion and keyhole porosity), and to generate processing maps for three novel aluminum alloys (AlFeCrSi, AlFeCrTi, and AlFeCrSiTi) designed for Laser-Powder Bed Fusion (LPBF). A melt pool database of 1386 samples was first generated using FEM simulations incorporating laser and material parameters, and used to train an initial ANN model. The model was then fine-tuned via transfer learning with experimental data from 108 additively manufactured samples of the three alloys. By leveraging simulation data, the TL-ANN achieved high predictive accuracy with minimal experimental input, enabling efficient process optimization and defect mitigation in LPBF. While the trained source model is transferable to other aluminum alloys with limited data, the proposed methodology is also applicable to other metallic materials, allowing rapid and cost-effective generation of processing maps. The present work comprises 3 documents: - 1 MS Word document explaining how to use the MS Excel document and the Python script provided (the present document). - 1 MS Excel document including a database of melt pool dimensions obtained by 1386 FEM simulations on three Al-Fe-Cr-Ti, Al-Fe-Cr-Si, Al-Fe-Cr-Si-Ti LPBF alloys (considering various processing parameters) and predicted defects (LoF and Keyhole) assuming Hatch Distance (HD), Point Distance (PD) and Layer Thickness (LT) of 130, 80 and 25 µm, respectively. Additionally, 108 experimental data of printed samples used to calibrate the FEM data using the machine learning algorithms given in the associated python script. This MS Excel document can be modified to include or substitute the readers experimental data. - 1 Python script to run ML and TL-ML on the given MS Excel document, giving melt pool sizes and processing maps including LoF and Keyhole defect regions for a wide range of laser speed vs. laser power processing conditions. The reader can use this script to predict processing maps from new experimental data by including or replacing them on the given MS Excel document. The new alloys can be different from aluminium ones. The best performance is expected for melt pools approximately semi-spherical. Although the MS Excel file includes data from 3 experimental alloys, the processing maps (figures) obtained from the script are given initially for one of them. Thus, the python script, in its present version, should be slightly modified by hand, choosing, at lines 599-601, one of the three materials. The python script can be adapted easily to other situations.