Three-dimensional U-Net models can use XCT volume data to identify internal defects in additive manufacturing parts. A study comparing three model variants reports mean intersection over union scores of 0.863 to 0.884, while highlighting practical limits around training time, defect boundaries, and generalization.