Abstract
Additive manufacturing is a rapidly growing industry with significant growth over the past decade. As the technology continues to evolve, there are many opportunities for growth and innovation in the additive manufacturing industry. It is exceedingly challenging to control all process parameters or assess the combined impact of all the variables on the characteristics of a manufactured part. The creation of a functional object is manufacturing's primary objective. Advanced research related to machine learning concerning additive manufacturing and analysis of process parameters is becoming an urgent necessity.
This study aimed to analyse the laser additive manufacturing of titanium alloy using artificial neural networks, and several objectives were investigated. The first was building ANN models that relate SLM process parameters with six inputs and three measured properties. The second objective is to compare and determine the effectiveness of different ANN algorithms in predicting the SLM mechanical properties of titanium alloy. The third objective was to conduct a sensitivity analysis of ANN models, specifically focusing on the influence of varying the number of neurons in the hidden layer. The MATLAB platform developed, trained, and validated the neural network models. The study utilized datasets collected from existing literature on the topic. The data was normalised and smoothed before training with MATLAB.
From the models trained with the Levenberg Marquardt algorithm, the regression was as follows: yield strength = 0.99983, ultimate tensile strength = 0.9979, and elongation = 0.98896. The ANN model was trained with the Bayesian regularization and Levenberg Marquardt algorithm to predict yield strength. However, the best model was obtained with the Bayesian regularization algorithm. The MSE obtained was 4.5928×10-16 with an R correlation coefficient of 0.99998. The ANN model for the prediction of ultimate tensile strength produced the best results for MSE with 4.2819×10-16 and a regression value of 0.9998 when training with Bayesian. The results of the ANN model for predicting elongation had an MSE value of 1.3689×10-16 and a correlation coefficient value of 0.9997.
The results revealed that Bayesian regularization consistently outperformed Levenberg Marquardt across all measured properties, showcasing its superiority in terms of Mean Squared Error (MSE) and regression values. The high regression values show that ANN models can
v
efficiently and accurately predict the mechanical properties of a Titanium alloy during laser additive manufacturing. The variation in the number of neurons in the hidden layer unveiled crucial nuances in the ANN models. This insight contributes to refining the neural network architecture, maximizing its performance, and reinforcing the reliability of predictions.
The study's results provide valuable insights into the complex, nonlinear relationships between the SLM process parameters and the resulting mechanical properties of the titanium alloy, contributing to the development of more accurate and reliable models for the SLM process parameters of titanium alloy. The validated ANN models, particularly those employing Bayesian regularization, are valuable tools for predicting mechanical properties and facilitating informed decision-making in additive manufacturing.
Keywords: Artificial Neural Network, Bayesian Regularization, Levenberg Marquardt,
Predictions, Selective Laser Melting, Titanium alloy.