Thursday, August 22, 2019

A New Machine Learning for Additive Manufacturing (ML4AM) Project on Shape Deviation Generator and Learner


A new machine learning for AM (ML4AM) project, "NSF CMMI-1901514: Shape Deviation Generator and Learner- An Engineering-Informed Convolution Modeling andLearning Framework for Additive Manufacturing Accuracy Control", has been funded by US National Science Foundation. (PI: Huang, $350K, 08/2019∼ 07/2022).  The grant abstract is given below.



Although additive manufacturing (AM), known as 3D Printing, holds great promise as a direct manufacturing technology, significant deviations from the desired part shape often occur in the printed parts.  As a result, shape distortion control is a critical issue for AM built products. With advances in computing and increased accessibility of AM product data, machine learning for AM has become a viable strategy for enhancing 3D printing performance. However, meaningful learning of engineering data requires effective integration of domain knowledge, making general-purpose machine learning methods difficult to apply. As a result, there is a critical need for an engineering-informed, data-analytical, machine learning framework for shape distortion control.  Such a tool is essential to improving AM quality and reducing cost and waste.



The project will establish an engineering-informed convolution modeling and learning methodological framework for AM distortion control. A Shape Deviation Generator and Learner for shape accuracy control will be researched by: (1) modeling 3D shape deviation generation by establishing a new convolution formulation for layer-by-layer fabrication processes, (2) transferring the 3D shape deviation model from a small set of training shapes to a wider variety of shapes by exploring and learning shape similarity under a cookie-cutter modeling framework, (3) transferring the shape deviation model between AM processes by exploring and learning process similarity through an effect equivalence framework, and (4) validating modeling and transfer learning methodologies in both polymer and metal AM processes. Methodologies and tools will be developed to mitigate both shape and process complexities towards the goal of building AM products with high geometric fidelity: from single shape to multiple shapes, and from single process to multiple AM processes.