Material extrusion 3D printing has long been established for rapid prototyping and functional testing in many research and industry fields. However, its inconsistency and intrinsic defects (surface roughness and geometric inaccuracies) hinder its application in several areas, most notably “certify-as-you-build” small-batch prototyping and large-batch production. In this study, we present an approach to reduce both inconsistency and the 3D geometric inaccuracies of products fabricated by material extrusion. To achieve these improvements in print quality, we developed an in situ metrology system, which scans each layer at the time of printing, providing a 3D model of the as-printed part. We then trained machine learning algorithms with data from this scanning system and predicted 3D geometric inaccuracies in new designs. Eight conditional adversarial network (CAN) machine learning models were trained on a limited number of scanned profile images of different layers, consisting of less than 50 actual images and 50 generated images, to predict the 3D geometric deviations of freeform shapes. The generated images were produced by randomly combining and cropping the actual images without any distortion. These CAN models produced predictions where at least 44.4%, 87.6%, 99.2% of data were within ± 0.05 mm, ± 0.10 mm, ± 0.15 mm of the actual measured value, respectively. A laser sensor was integrated into a material extrusion 3D printer to achieve in situ monitoring of dimensional inaccuracies during printing, which leaves the door open to implement a closed-loop feedback system to compensate geometric errors during printing in the future and fabricate “certify-as-you-build” products.

长期以来，材料挤压3D打印已经在许多研究和工业领域中用于快速原型制作和功能测试。但是，它的不一致和固有缺陷（表面粗糙度和几何精度）阻碍了它在多个领域中的应用，尤其是“按需认证”的小批量原型设计和大批量生产。在这项研究中，我们提出了一种减少材料挤压制造的产品的不一致性和3D几何误差的方法。为了实现打印质量的这些改进，我们开发了一种原位计量系统，该系统可在打印时扫描每一层，从而提供已打印零件的3D模型。然后，我们使用来自该扫描系统的数据训练了机器学习算法，并预测了新设计中的3D几何误差。八个条件对抗网络（CAN）机器学习模型在有限数量的不同层的已扫描轮廓图像上进行训练，这些图像由少于50个实际图像和50个生成的图像组成，以预测自由形状的3D几何偏差。生成的图像是通过随机组合并裁剪实际图像而产生的，没有任何失真。这些CAN模型产生了预测，其中至少44.4％，87.6％，99.2％的数据分别在实际测量值的±0.05 mm，±0.10 mm和±0.15 mm范围内。激光传感器已集成到材料挤出3D打印机中，以实现在打印过程中原位监控尺寸误差，