Additive Manufacturing

AM processes of laser powder bed fusion (LPBF) and fused filament fabrication (FFF) are modeled. The different physics modeled are heat transfer, structural deformation, phase transformation, sintering, etc.

Computational techniques used:

• Analytical models
• Finite element analysis (Abaqus)
• Cellular automata
• Computational fluid dynamics (OpenFOAM)

Quantities of interest of the additively manufactured part:

• Melt pool size
• Part geometry
• Residual stress
• Sintering neck growth
• Microstructure
• Porosity/Voids

Experiments for fused filament fabrication AM process are conducted on desktop printing machine. The quality of the printed parts is measured using:

• Digital caliper
• Laser displacement sensor / profilometer
• Infrared thermal camera
• Microscopy

The experiments are used to quantify the error in model prediction, and also to validate the results from model prediction-based process design and control.

Physics-based models for AM are computationally intensive. The physical experiments for AM, especially metal AM, are also expensive due the high cost of raw materials. To overcome these challenges, we build machine learning (ML)-based surrogate models, using limited data from physics-based models and physical experiments. We use the ML models to support uncertainty quantification, process optimization, and process control.

ML models pursued:

1. Gaussian process (GP) models
2. Deep neural networks (DNN)

Physics-Informed Machine Learning (PIML)

Physics-based models are limited by their computational complexity or incomplete physics. In contrast, machine learning models appear promising for complex systems, given adequate quality and quantity of data. The combined use of physics-based and machine learning models achieves more accurate and physically consistent predictions by leveraging the advantages of each method. Multiple PIML strategies and their combinations are investigated to improve the prediction accuracy and physics consistency of machine learning models.

Variability is a significant challenge in AM. Since AM experiments are expensive, decisions such as process optimization and control cannot be achieved using trial-and-error experiments; thus a model-based approach is necessary. However, multiple sources of uncertainty such as input uncertainty, model parameter uncertainty, model errors, etc. propagate through multiple computational models and cause uncertainty in the quantity of interest. Our effort has developed a systematic methodology for uncertainty quantification, by properly aggregating the contributions of multiple types of uncertainty sources. Global sensitivity analysis is used to quantify the contribution of each uncertainty source to the uncertainty in quantity of interest. These results are utilized for process uncertainty management, and resource allocation for uncertainty reduction.

Model prediction corrected using experimental data is used for determining optimal process parameters in fused filament fabrication (FFF), instead of a physical experiments-based trial-and-error approach. Methods for optimization under uncertainty, such as reliability-based design optimization (RBDO), and robustness-based optimization (RBO) are studied.

Parts printed with optimal process parameters such as nozzle temperature, nozzle velocity, and layer thickness are found to have better geometric accuracy and mechanical properties, and also less voids and porosity.