Data-driven strategies for control in additive manufacturing

Thumbnail Image
Inyang-Udoh, Uduak
Issue Date
Electronic thesis
Mechanical engineering
Research Projects
Organizational Units
Journal Issue
Alternative Title
Additive Manufacturing (AM) techniques are quickly becoming attractive for fabricating parts ranging from biological tissues to aircraft components. A key challenge in these techniques is controlling the quality of parts to ensure high accuracy, throughput and repeatability. Given the fast and high-dimensional multi-scale dynamics associated with many AM processes, it is often difficult to physically model these processes for control. This thesis examines how data obtained from the process may be used to implement feed-forward and feedback control. Two AM processes are studied: Inkjet 3D Printing, commonly used for fabricating polymer parts; and Selective Laser Melting (SLM) used for producing metal parts. Inkjet 3D printing builds, or prints, 3D parts with high precision by sequentially depositing and hardening liquid material. Because of the complex fluid dynamics at play and multiple printing parameters to tune, the dynamic process is challenging to model. There is significant prior work on modeling local droplet behavior, without necessarily considering the overall part-level dynamics. On the other hand, studies concerned with part geometry control have developed reduced order models that do not well capture nonlinear fluid behavior. In this research, to implicitly learn the fluid dynamics for geometry-level control, we employ machine learning strategies but push their conventional usage in the following ways: (1) We use a shallow multi-layer neural network to capture the geometrical relationship between printing input pattern and measured output profiles. (2) We develop a data-driven physics-guided model that uses a recurrent neural network to model droplet spreading and coalescence. Not only does this model capture the complex fluid behavior, it is formulated to do so at the geometry level. We validate the model on data collected from an inkjet 3D printing setup. The proposed model outperforms a blackbox off-the-shelf multilayer perceptron (neural network) by using significantly less data for training, at the same time delivering better performance in RMS error on test data. The proposed model is also compared with a state-of-the-art reduced order linear model and shows substantial improvement in RMS error on test data. Experimental results also underline that the model parameters learned are geometry invariant, that is, the model parameters trained on one geometry can be used to predict the height map evolution for other geometries without relearning. Next, we propose and demonstrate a new predictive control algorithm that leverages the neural-network-like structure of the model. Back-propagation is used for efficient gradient calculations to determine optimal control inputs, namely droplet patterns for subsequent layer(s), to optimize a quadratic cost function. Further, we analyse the stability of the open-loop and closed-loop printing system based on the developed model and control scheme. In addition, we develop an efficient algorithm to make online learning and feedback control practically feasible purposes. Simulation and experimental results show that both feedforward and feedback control substantially improve the height profile over existing control approaches. In line with the data-driven theme of this thesis, we investigate a data-driven approach for controlling a metal AM process, selective laser melting (SLM), where a laser beam is to local melt metal powder and create complex geometries in a layer-by-layer manner. Similar to the inkjet process, because of the complexity associated with the heating, melting, cooling and solidification, it is difficult to model the SLM process in a control-oriented fashion. To address this, this thesis presents a model-free iterative learning control (ILC) scheme for designing laser power profiles for multi-objective temperature control in SLM. The goal is to ensure that while temperature distribution in the selected region is sufficient to cause melting and fusion, the meltpool is not overheated. We first formulate this goal as an optimization problem with the power profile as the decision variable and the cost function to be minimized being the sum of two unidirectional error terms (for upper and lower temperature bounds, respectively). Given the difficulty in analytically modeling the temperature-laser power relationship in SLM for gradient computations as in standard ILC, we solve the minimization problem using a model-free ILC scheme. In this scheme, the control input that minimizes the cost function is learned through a data-driven gradient descent update that uses the process itself to compute the gradient direction. The gradient descent algorithm proposed here accounts for the time-varying behavior of the SLM thermal dynamics because of the scan path. This is accomplished by feeding the temperature output error, reversed in time, through the process itself with a reversed scan path direction. For validation, this multi-objective gradient-based ILC algorithm is implemented on a three-phase high-fidelity simulation of the SLM process. The results demonstrate the algorithm's ability to drive the temperature distribution to within a prescribed range in scenarios where standard (single-objective constant gain) ILC fails.
December 2021
School of Engineering
Full Citation
Rensselaer Polytechnic Institute, Troy, NY
Terms of Use
PubMed ID