Modelling of geometry for directed energy deposition via machine learning

Abstract

Predicting the geometry of bead in multi-track and multi-layer Directed Energy Deposition (DED) presents a challenge due to variations in process parameters, resulting in changes in geometry from one layer or track to another. To address this issue, a Long Short-Term Memory (LSTM) model is applied to model and predict the bead geometry in multi-track and multi-layer deposition, specifically height and waviness by predicting the 2nd order polynomial coefficients. The LSTM model is designed to capture long-term dependencies in the dataset containing process parameters and corresponding bead geometries. The process parameters used in this study include laser power, scanning speed, powder mass flow rate, hatch spacing value and z-incremental value. The applied model is trained, validated and tested using process parameters and the corresponding coefficients of unique sample name and layer number combination.

The LSTM model achieved a considerable good 𝑅2 score of 0.81 for the coefficients, and 0.77 for height, but a poor score of -0.79 for waviness. The model’s performance in predicting the detailed waviness variations between the curves requires further improvement. Despite the limitations in waviness prediction, the model demonstrated its usefulness in capturing general trend of bead geometry changes under varying process parameters, particularly laser power and powder mass flow rate. To further assess the model, standard deviation of the curve heights are also analysed through which it revealed that adjusting the aforementioned parameters has the potential to decrease the variability in the curves height. This study highlights the potential of using a many to many LSTM model in bead geometry modelling and parameter selection in the field of additive manufacturing.