Session: 17-01-01: Research Posters

Paper Number: 149966

149966 - Integrated Fem and Statistical Learning Approach for Characterizing the Effect of Process Parameters on Material Design 

Additive manufacturing is a useful manufacturing technique that allows for complex designs that is especially useful for prototyping. However, the interplay between the final results and the process parameters of additive manufacturing is very intricate. Thermomechanical simulation, in conjunction with statistical and machine learning, can be an effective design tool to guide manufacturing parameters for material designs. This is particularly beneficial when the material used is expensive or time intensive to manufacture. In this work, we demonstrate the applicability of the combined approach of thermomechanical simulations and statistical/machine learning for additive manufacturing metal alloys. Specifically, in this work, we use data from 192 thermomechanical finite element-based simulations performed on nickel alloys to quantify stress and deformation as a function of a range of additive manufacturing process parameters such as laser power, laser scan speed, layer height, and various cooling assumptions. The geometry and material was kept fixed between simulations. Nodes from the thermomechanical simulation were averaged with neighboring nodes to allow for high-fidelity simulations while reducing the size of the dataset used in statistical/machine learning. Various levels of averaging were initially considered but the results between the levels were similar so only one level is shown.

 

We utilize unsupervised learning methods with statistical insights to quantify the relationship between the process parameter settings and the outcome variables. The finite mixture model, known for its flexibility, was utilized to better capture the variability while remaining interpretable. In the first approach, a mixture of Gaussians is fitted to cluster the von Mises stress and the three spatial displacement components and then regression models, using the von Mises stress or deformation as the response with the process parameters being the regressors, are fit within each cluster. In the second approach, clustering and regression analysis are done simultaneously through finite mixtures of regression models. The two models are compared along with a naive baseline that does not consider any clustering, and the second approach resulted in improved performance and identification of potential design weak points. Our finding predicts overlapping process parameter settings that lead to optimum outcomes. Different sections of the geometry had different reactions to the changes in each process parameter. The relative change in overall von Mises stress and deformation caused by increasing the laser power was greater on sections that are relatively unsupported when compared to portions that are more stable. The results were similar for laser scan speed but increasing the scan speed reduced the expected von Mises stress and deformation. 

 

Presenting Author: Jason Hasse South Dakota State University

Presenting Author Biography: Jason is a graduate student at South Dakota State University pursing a Ph.D. in Computational Science and Statistics. His primary research interest is in finite mixture models with applications in materials science, health care, and public health.

Authors:

Jason Hasse South Dakota State University
Semhar Michael South Dakota State University
Anamika Prasad Florida International University