Predictions of in-situ melt pool geometric signatures via machine learning techniques for laser metal deposition

Laser metal deposition (LMD) can produce near-net-shape components at high build-up rates for many applications, e.g. turbine blades, aerospace engine parts, and patient-specific implants. However, builds suffer from distortion and defects associated with ineffective process control. For example, melt pool features including height, depth, and dilution are transient, while process parameters including laser power, scanning speed, and powder feed rate remain constant in an open-loop LMD system. Improving product quality requires estimating these transient features to enable process control. This paper presents a semi-dynamic, data-driven framework to address this challenge. The framework correlates combined process parameters (laser power, scanning speed, powder feed rate, line energy density, specific energy density) and features from melt pool thermal images (melt pool width, area, mean temperature, maximum temperature) with hard-to-monitor, melt-pool-related features (height, depth, dilution). Sixty single-track experiments were conducted to acquire sensing data and dimensions of the track cross-sections. Significant input features for training machine learning (ML) models were selected based on Spearman's rank correlation coefficient. Results show that the correlation between hard-to-monitor melt-pool-wise features, combined process parameters, and limited in-situ sensing data are described well by the models presented here. Critically, an artificial neural network (ANN) showed the best performance.