Session: 03-08-02: Computational Modeling and Simulation for Advanced Manufacturing

Paper Number: 139424

139424 - An Alternative Machine Learning-Based Approach to Establish Stronger Correlations Between Laser Powder Bed Fusion (Lpbf) Processing Parameters and Part Properties 

Laser powder bed fusion (LPBF) is a layer-wise form of metal additive manufacturing (AM) that has gained considerable popularity in various sectors due to its dazzling ability to fabricate complex geometries, fast prototyping, cost-effectiveness, design freedom, waste minimization, and manufacturability of technologically essential metals such as nickel, copper, aluminum, titanium, etc. The processing parameters have been proven to affect the quality and performance of the final products, such as porosity content, hardness, strength, cracks, etc. The four most widely researched major processing parameters include laser powder ( ), scanning speed ( ), hatching space ( ), and layer height ( ). Due to the involvement of complex Multiphysics phenomena during the LPBF process, it is an arduous task to establish clear and meaningful correlations between these processing parameters and the final product properties or performance metrics. Hence, there has been an enormous research motivation to reveal and understand all such possible correlations. Although several researchers have developed and studied some numerical simulations, the prediction of part properties or performance metrics has yet to become reliably accurate and offers acceptable margins of error. Numerical simulations can be computationally intensive, slow, expensive, and complicated in view of error traceability and uncertainty estimations. Therefore, there has been a rapidly growing interest in applying state-of-the-art machine learning (ML) algorithms for modeling and predicting the performance or properties of AM parts, taking advantage of their extraordinary ability to map the hidden relationships of input and output parameters. Bearing in mind that the processing parameters are on substantially different numerical scales and physically different (e.g.,  = 0.08 mm vs.  = 1250 mm/s), simple mathematical normalization strategies might achieve high accuracy but lack a tangible physical meaning. In contrast, this paper aims to utilize a widely accepted physical normalization method along with two sets of dimensionless numbers to generate better input parameters for the considered ML algorithm. For the LPBF processing of IN 625, this approach led to delineating stronger correlations between the normalized parameters and the hardness of AM parts. Following this approach, the R-squared value increased to ~92%, and RMSE dropped to ~0.3.

Presenting Author: Feiyang Bai University of the DC

Presenting Author Biography: Feiyang is a phd candidate of Mechanical Engineering at University of the DC. He is working on developing machine learning based model to simulate and understand the process involved in AM and materials used in AM.

Authors:

Feiyang Bai University of the DC
Faridreza Attarzadeh University of the DC
Jiajun Xu University of the DC