Powder-based Laser Metal Deposition (LMD) offers a promising additive manufacturing process given the large number of available materials for cladding or generative applications. In laser cladding of dissimilar materials, it is necessary to control the mixing of substrate and additive in the interaction zone to ensure safe metallurgical bonding whilst avoiding critical chemical compositions that lead to undesired phase precipitation. However, the generation of empirical data for LMD process development is very challenging and time-consuming. In this context, different machine learning models are examined to identify whether they can converge with a small amount of empirical data.

In this work, the prediction accuracy of Back Propagation Neural Network (BPNN), Long Short-Term Memory (LSTM) and eXtreme Gradient Boosting (XGBoost) were compared using Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE). A hyperparameter optimization was performed for each model. The materials used are 316L as substrate and VDM Alloy 780 as additive. The data set used consists of 40 empirically determined values. The input parameters are laser power, feed rate and powder mass flow rate. The quality characteristics height, width, dilution, Fe-amount and the seam contour are defined as output.

As results, the predictions were compared with retained validation data and described as MSE and MAPE to determine the prediction accuracy for the models. BPNN achieved a prediction accuracy of 0.0072 MSE and 4.37% MAPE and XGBoost of 0.0084 MSE and 6.34% MAPE. The most accurate prediction was achieved by LSTM with 0.0053 MSE and 3.75% MAPE.