Printability for additive manufacturing with machine learning: Hybrid intelligent Gaussian process surrogate-based neural network model for Co-Cr alloy


Mahmood M. A. , Rehman A. U. , Karakas B., Sever A., Rehman R. U. , SALAMCI M. U. , ...More

JOURNAL OF THE MECHANICAL BEHAVIOR OF BIOMEDICAL MATERIALS, vol.135, 2022 (SCI-Expanded) identifier identifier identifier

  • Publication Type: Article / Article
  • Volume: 135
  • Publication Date: 2022
  • Doi Number: 10.1016/j.jmbbm.2022.105428
  • Journal Name: JOURNAL OF THE MECHANICAL BEHAVIOR OF BIOMEDICAL MATERIALS
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Biotechnology Research Abstracts, Compendex, EMBASE, INSPEC, MEDLINE, Metadex
  • Keywords: LPBF, Co-Cr, Stress -strain curve, Process -property relationship, Experiments, Gaussian surrogate model, Machine learning, METAL-DEPOSITION, MICROSTRUCTURE, OPTIMIZATION, COMPONENTS, TI-6AL-4V, SCIENCE, DESIGN, STEEL
  • Gazi University Affiliated: Yes

Abstract

AM has revolutionized the manufacturing industry, involving several operating parameters that may affect the properties of the final manufactured part. In AM, LPBF has proved its reliability in producing dense components; however, process development for every material necessitates extensive testing. Even the tiniest change can negate all the data for the same material. It is vital to have a P-P correlation that can train itself following a change in powder or machine to achieve defects-free parts and optimal properties. These goals cannot be met alone by multi-physics. One of the ways to address this issue is to apply ML, but it requires a huge data set for training and testing purposes. A framework has been developed for Co-Cr S-S curves to resolve this issue. Twenty-two experimental S-S curves have been generated to produce YS, TS, and EL data points. In combination with DNN, these data points have been applied to the validated and tested GPS-surrogate model to develop a smart processing window to achieve desired YS, TS, and EL. LP, LSS, HD, and PLT have been selected during the whole framework as inputs, while YS, TS, and EL have been classified as outputs. The output of the smart window was verified experimentally. It is found that the highest YS (1110.91 MPa) is attained using LP = 180 W, LSS = 600 mm/s and HD = 70 mu m, while least YS (645.05 MPa) is identified using LP = 160 W, LSS = 900 mm/s and HD = 70 mu m. For TS, the maximum (165.91 MPa) and minimum (689.73 MPa) values have been achieved using LP = 180 W, LSS = 900 mm/s and HD = 70 mu m, and LP = 180 W, LSS = 1000 mm/s and HD = 70 mu m, respectively. In the case of EL, LP = 180 W, LSS = 700 mm/s and HD = 70 mu m, and LP = 180 W, LSS = 600 mm/s and HD = 70 mu m, resulted 23.04% and 0.789% EL, respectively. Using CC, LP and HD did not significantly affect the TS, YS, and EL, while a negative relationship has been found for LSS with TS, YS, and EL. The smart processing window showed that the YS and TS could be achieved at low-high LP and low LSS at the cost of EL. This study provides a technique for framework development in the case of P-P relation based on the provided inputs and the corresponding outputs, leading toward process smartification.