This dissertation focuses on improving part quality and process reliability by proposing methods for predicting process outcomes, detecting defects and process anomalies in DED. These methods include physics-based process modeling, in-situ process monitoring, statistical and machine learning algorithms. Furthermore, these methods are utilized to assist the post-process or layer-intermittent machining as a corrective action in response to defects or process anomalies.

The physics-based model predicts the melt pool temperature, keeps track of the thermal history, and simulates the deposition geometry using a voxel-based approach by discretizing the scan path. This approach provides the capability of simulating 3D objects, where the layer-based scan path includes not only 1D deposition tracks, but also 2D features such as curvatures.

In-situ vision data acquisition, feature extraction, and analysis are performed to propose a method for detecting target regions in the laser-material interaction zone based on a low-cost high-dynamic-range (HDR) vision sensor. Adaptive image thresholding, connected component analysis, and iterative energy minimization are used to identify target regions in the field of view. The method is designed to be adaptive, in terms of obtaining parameters based on simple training data, and robust, in terms of feature detection performance subject to under-melt, conduction and keyhole melting mode phenomena.

The performance of the proposed region detection scheme is quantitatively and qualitatively evaluated against annotated data. It was found that the True Positive Rate in detection was above 90%, while the False Detection Rate was less than 10%.

The proposed feature extraction algorithm from melt pool images and the physics-based modeling are then leveraged to define, identify and classify regions of process stability in DED. The research efforts are focused on generating process maps to identify unstable process zones, with a reference to process physics, process signatures, and process outcomes using analytical modeling, in-situ melt pool monitoring, and ex-situ characterization, respectively. The goal is to classify the process signatures in pre-defined process zones (under-melt, conduction, keyhole, balling) to avoid instabilities, defects and anomalies using a low-cost high-dynamic range camera and a kNN (k-nearest neighbor) classifier, which has achieved 13\% error rate. With this approach, decisions may be made to perform corrective actions (e.g. machining, re-manufacturing), or to scrap the manufactured part.

A dual wavelength pyrometer was also deployed to monitor the laser-material interaction zone coaxially with the laser beam to validate the proposed physics-based thermomechanical model in a multi-layer, multi-track 3D part manufactured via DED process. The pyrometry data was further analyzed to find a correlation between the melt pool signatures and the process outcomes and to detect geometric defects. Such defects can be addressed with post-process machining via two correction strategies. In the first approach, machining is used as a post-processing operation, where geometric fidelity of the overall part is not met and/or surface quality requires improvement. The second strategy is a layer-intermittent approach, where the deposition quality of each layer is assessed and layers are flagged for machining based on the melt pool signatures.

This dissertation contributes towards accelerating the industrial adoption of DED and advancing the competitive advantages of metal additive manufacturing compared to conventional manufacturing processes. This is achieved by improving the part quality, process performance and reliability by advancing the DED process modeling, process monitoring and data analysis, and by proposing an intelligent decision-making schema for a hybrid machining approach.