An impulse response formulation for small-sample learning and control of additive manufacturing quality

Machine learning for additive manufacturing (ML4AM) has emerged as a viable strategy in recent years to enhance 3D printing performance. However, the amount of data required for model training and the lack of ability to infer AM process insights can be serious barriers for black-box learning methods. Due to the nature of low-volume fabrication of infinite product variety in AM, ML4AM also faces “small data, big tasks” challenges to learn heterogeneous point cloud data and control the quality of new designs. To address these challenges, this work establishes an impulse response formulation of layer-wise AM processes to relate design inputs with the deformed final products. To enable prescriptive learning from a small sample of printed parts with different 3D shapes, we develop a fabrication-aware input–output representation, where each product is constructed by a large amount of basic shap primitives. The impulse response model depicts how the 2D shape primitives (circular sectors, line segments, and corner segments) in each layer are stacked up to become final 3D shape primitives. A geometric quality of a new design can therefore be predicted through the construction of learned shape primitives. Essentially, the small-sample learning of printed products is transformed into a large-sample learning of printed shape primitives under the impulse response formulation of AM. This fabrication-aware formulation builds the foundation for applying well-established control theory to the intelligent quality control in AM. It not only provides theoretical underpinning and justification of our previous work, but also enable new opportunities in ML4AM. As an example, it leads to transfer function characterization of AM processes to uncover process insights. It also provides block-diagram representation of AM processes to design and optimize the control of AM quality.