A novel feature engineering approach for predicting melt pool depth during LPBF by machine learning models

• Published in: Additive manufacturing letters Vol. 10; p. 100214
• Main Authors: Mosallanejad, Mohammad Hossein, Gashmard, Hassan, Javanbakht, Mahdi, Niroumand, Behzad, Saboori, Abdollah
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.07.2024; Elsevier
• Subjects: Additive manufacturing; Artificial intelligence; Atomic features; Machine learning; Melt pool depth; Rosenthal equation; Rosenthal equation; Additive manufacturing; Atomic features; Artificial intelligence; Melt pool depth; Machine learning
• ISSN: 2772-3690; 2772-3690
• DOI: 10.1016/j.addlet.2024.100214

•Machine learning algorithms were trained to develop models predicting the melt pool depth during the LPBF procedure. High-fidelity data was collected from the literature to ensure the training quality of the investigated machine learning models.•A physics-informed feature selection was adopted to enhance the prediction capability of the trained models. For the first time, new atomic features of the studied commercial alloys were generated and considered for training machine learning models to predict the melt pool depth during the LPBF process. The atomic features, closely tied to the electronic structure and bonding attributes of the material, which play a pivotal role in determining its thermal and physical characteristics, were expected to account for more factors that affected the melt-pool depth.•The prediction accuracy of best-performing machine learning models was compared with that of the Rosenthal equation using a new dataset not seen by the ML models. The XGBoost model was found to outperform the Rosenthal equation in terms of R2. Melt pool geometry is a deterministic factor affecting the characteristics of metal Additive Manufacturing (AM) components. The wide array of physical and thermal phenomena involved during the formation of the AM melt pool, along with the great variety of alloy compositions and AM methods, coupled with the clear influence of multiple process parameters, make it difficult to predict the melt pool geometry under a given set of conditions. Therefore, using Artificial Intelligence (AI) approaches such as Machine Learning (ML) is necessary for accurate predictions. Using a physics-informed feature selection strategy along with the application of atomic features for the first time, this work aims to offer accurately trained models relying on existing high-fidelity data for most common alloys in AM academia and industry, i.e., 316 L stainless steel, Ti6Al4V, and AlSi10Mg. Multiple ML algorithms were trained, and the results revealed that the average R2 and RMSE obtained by the K-fold cross-validation (K = 5) were significantly enhanced when laser and material properties, inspired by the analytical models for AM melt pool geometry, were used as the model features. Removing the excess features and applying atomic features further enhanced the accuracy of the models. As a result, R2 for the XGBoost, CatBoost, and GPR models were 0.907, 0.889, and 0.882, respectively, while the hold-out cross-validation led to 0.978, 0.976, and 0.945, respectively. Furthermore, the results showed that the XGBoost model outperforms the Rosenthal equation. This approach provides a pathway to more accurately predict the properties of metal AM components.