PSI - Issue 34

N. Lammens et al. / Procedia Structural Integrity 34 (2021) 247–252 N. Lammens/ Structural Integrity Procedia 00 (2019) 000 – 000

250

4

State-of-the-art methods in high-fidelity modelling of microstructural formation and evolution (e.g. phase field based models) are currently still too computationally intensive to be feasibly applied to realistic component sizes. However, models based on more simple methods such as cellular automata and phase transformation equations allow a faster estimate of microstructural properties at the meso-level. Figure 4 (left) shows how starting from a meso-scale temperature prediction, phase fractions are predicted for a Ti 6Al-4V material. Figure 4 (right) shows a solidification simulation for IN718 accounting for scanning strategy and the resulting effect on grain structure. Such simulation tools allow to assess the impact of the AM process on the local material properties enabling thus the first link in the P3 concept: from process to property.

Figure 4. (left) Phase fraction prediction on a meso-scale model for Ti-6Al-4V during an AM process. (right) Solidification grain structure predicted for a bi-directional scanning strategy of an IN718 sample 4. Performance simulation The next step in the P3 concept is to predict the impact of the AM process induced different local material properties on the performance of the 3D printed component. 4.1. Machine Learning fatigue model To predict fatigue performance of AM components, a material model based on machine learning is adopted (Lammens et al. (2019)). This technique allows to account for complex interactions between many different influencing parameters such as surface roughness, build orientation, and surface and heat treatments. Further parameters can be added to the model, provided test data is available to train the model. Figure 5 shows the experimental data obtained on simple test coupons in the current work and compared to the machine learning model trained on previous test data (obtained on a different printer).

Figure 5. Experimental fatigue data and predicted performance using the machine learning implementation