Abstract The Direct Energy Deposition (DED) process is a metal additive manufacturing (AM) technique that uses high-energy heat sources, such as lasers, to melt metal materials, enabling the production of three-dimensional (3D) structures. Due to its highly flexible deposition head, this process is primarily used for part repairs, allowing localized repair of damaged areas without replacing entire components. These advantages make DED particularly valuable in high-value industries such as aerospace and automotive. Given these applications, the DED process must ensure precise and rapid handling of targeted areas. However, this process, which is performed by melting and solidifying metal, may result in quality differences due to the influence of external factors such as temperature and humidity and process parameters such as laser power and material feed density. Additionally, since the process is layered, the initial unevenness may accumulate, degrading the overall structure’s geometrical quality. This geometrical irregularity can further exacerbate instability in heat and material supply, potentially leading to internal defects like pores or lack of fusion. These defects can critically impact part quality in high-value industries, which operate in high-temperature and high-pressure environments, weakening mechanical properties and compromising both safety and performance. Therefore, real-time geometry monitoring is essential to ensure the stability and reliability of the DED process. This enables rapid detection and correction of irregularities, supporting economical, high-quality additive manufacturing. To establish such a geometry monitoring system, this study developed an artificial intelligence (AI)-based height prediction model using image and temperature data of the melt pool acquired during the process. First, laser power and scanning speed, the parameters most influential to process quality, were selected as key process parameters, and a full factorial design with 2 factors and 3 levels was applied for experiments. Under these conditions, the melt-pool multi-sensor data were collected through non-contact sensors such as CCD and IR cameras, and the line scanner captured additional geometric information of the deposited structure. Collected data were pre-processed for noise removal and target information filtering. Following this, features related to the size, brightness, position, shape, and temperature of the melt-pool were extracted, with a subset of statistically significant features chosen for model training. In this experiment, various time sequences and data lengths appeared due to differences in tool path or process conditions. To solve this, data transformation techniques were proposed, and as a result, data could be learned under various conditions in a single model. This technique involved time-sequential alignment for each track and interpolation to match the data point lengths. Finally, several AI models were constructed for the height prediction using the melt-pool multi-sensor data. Among them, the one-dimensional convolutional neural network (1D-CNN) model, which showed the best mean absolute percentage error (MAPE) performance, selected as the final height prediction model. This foundation enables real-time geometric prediction and monitoring and can contribute to maintaining optimal quality throughout the DED process.
Shin et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: