Physics model-based probabilistic process optimization and control

Abstract

A model-based approach for process design and control is more efficient and economical than the physical experiment-based trial-and-error approach in advanced manufacturing processes such as additive manufacturing (AM). However, while employing predictive models it is necessary to take into account the various sources of uncertainty in the computational models in addition to the variability in the AM process. This work presents a Bayesian methodology for building a probabilistic digital twin that integrates physics-based modeling and experimental data, which is then deployed for process design and layer-wise process control in AM. The methodology consists of three steps: model development, initial optimization of process parameters, and adjustment of process parameters during manufacturing.

Model development: First, a finite element analysis (FEA) model is used to predict porosity in a laser powder bed fusion (LPBF) AM process. Since the FEA model is expensive, a faster surrogate model is constructed to replace the original prediction model and is used for process optimization and control. Data is acquired from in-situ monitoring (e.g., thermal camera) during manufacturing and ex-situ measurement (e.g., X-ray computed tomography) after manufacturing. The error between the model prediction and the observation is calibrated using the measurement data.

Initial optimization of process parameters: Based on the corrected prediction model and offline measurement data, the value of the initial process parameters to start the manufacturing process is determined. A robustness-based design optimization method is used to optimize the initial process parameter values while accounting for uncertainty.

Online process control: During manufacturing, the quality of the partially manufactured part is estimated based on the online process monitoring data. Since porosity is not directly observable during the printing process, the temperature profile obtained from the monitoring (using an infrared thermal camera) is used to infer the porosity in the partially finished part. The porosity inference model is constructed by first reducing the dimension of the thermal images by employing singular value decomposition. The prediction model is also updated at every layer based on the monitoring data, and the updated model is used to predict the porosity in the completed part. The predicted porosity is compared to the desired porosity and a decision is made whether the manufacturing should be terminated, continued as-is, or continued with modified process parameters. The effectiveness of the proposed method is demonstrated for controlling process parameters such as laser power and laser speed.