The process of laser metal deposition (LMD) is becoming increasingly popular due to its ability to manufacture intricate components for various industrial applications. However, ensuring the reliability of the deposition process has proven to be a challenging task. Despite the control mechanisms in place during the LMD process, process shifts may occur and negatively impact the component as deviations propagate along the layer growth direction. Relying solely on previous time steps to predict depositions on subsequent time steps may not be sufficient. To achieve robust real-time geometry control, this paper introduces a spatiotemporal adjacency-based predictive model that incorporates two types of features: (1) inter-neighbouring ( \(\zeta\) ) and (2) intra-neighbouring ( \(\xi\) ) features. The former samples the melt pool height of previous layers, whilst the latter includes the neighbouring heights of the predicted value within the same layer. To validate the proposed method, nine samples of bulk structures were fabricated. Initially, five different machine learning algorithms were compared without the adjacency features, from which long short-term memory was selected for downstream analysis. Three different sets of experiments were cross-validated to check the robustness of the selected predictive model. For the cuboid structure, in all three cases, with \(\zeta = 3\) and \(\xi = 1\) , the root-mean-square error (RMSE), mean absolute error (MAE), and R2 score of the model is the highest at 21.040 ± 0.251 µm, 16.339 ± 0.144 µm, and 0.719 ± 0.011, respectively. Nevertheless, experimental results have shown a significant improvement when the adjacent features are incorporated as additional input variables. Comparatively, when neither adjacent features are included, the RMSE, MAE, and R2 values are 29.984 ± 0.238 µm, 23.028 ± 0.100 µm, and 0.430 ± 0.032. These findings verify the effectiveness of the proposed method to increase the performance of the prediction model of the melt pool height.
