This research seeks to develop and test a framework for considering Inductive Hierarchical Analysis (IHA) using Machine Learning (ML) technique in a multistage material design of a gear manufacturing process. A ML model was developed, which implements Random Forest (RF) regression algorithm together with analysis of variance (ANOVA) approach for mapping sets of material and product design variables, thus classifying the design space into feasible and non-feasible solution space. This approach is applied to the design of steel gear within specified performance requirements by exploring the design space for the Process-Structure-Property-Performance (PSPP) relation in the hot rolling process. With an objective of how machine learning models can be incorporated into the material design and modeling for predicting the material properties of steel used for gear manufacturing, a working hypothesis was outlined, and tasks for accomplishing this objective was developed. Feasible solution space was derived by the RF approach across the PSPP chain for the gear design by considering forward and inverse mapping. For the forward process, feasible range set of 84.68 - 84.9 and 32.54 - 32.84 was mapped at the structural stage to give feasibility range of 193.169 - 193.1818 at the property stage and eventually a feasibility range of 1.578 - 1.676 for the performance stage. The feasibility range was used as an input for the inductive process to generate feasible region of 193.171 - 193.174 at the property stage, 32.61 - 32.69 and 84.41 - 84.70 at the structural stage, while a feasible range of 1340 - 1360K evolved at the processing stage. The inverse process shows propagated reduction in variable size across the PSPP linkage with a Root Mean Square Error value of . This indicates a negligible error propagation across the modeling analysis and shows the robustness of RF as an ensemble of multiple decision trees. To further test the developed model, a predictive analysis of the Time-Temperature Transformation (TTT) curve was carried out by implementing Monte Carlo simulations together with the RF algorithm. This method accurately predicts the TTT curve with an accuracy matrix of 92.7%. The application of Monte Carlo helps to increase the robustness of the RF model by generating more input data points.
This research work was extended to model the constitutive behavior of a flexinol wire to incorporate Artificial Intelligence (AI) with Finite Element Methods (FEM) as a way validating the approaching methods in this research. Being able to implement machine learning applications to computer integrated material design and modeling presents a novel approach to material design by using RF method. This broadens the knowledge base in material design processing and opens up opportunities for further research with the application of RF to material design. In addition, this research will help in bridging the gap and timeline in meeting customer’s performance requirements by serving as a tool for designing materials that are feasible with optimum performance capability for any engineering need.
Date of publication
Chung Hyun Goh, Ph.D., Fredericka Brown, Ph.D., Shih-Feng Chou, Ph.D.
M - Mechanical Engineering
Oyedele, Joseph O., "MACHINE LEARNING APPLICATIONS IN COMPUTER INTEGRATED MATERIAL DESIGN AND MODELING" (2019). Mechanical Engineering Theses. Paper 5.