Data-driven non-intrusive surrogate modeling and optimization of selective laser melting using machine learning methods

Abstract

The goal of the presented thesis is computational modeling and optimization of selective laser melting (SLM), which is an additive manufacturing(AM) process, using data-driven non-intrusive machine learning techniques. The selective laser melting process mostly depends on experiments to understand and predict the process. The use of traditional trial and error techniques to determine the AM process outputs could be time-consuming and costly. In addition, the numerical simulation takes a significantly longer time to compute. A new framework and techniques like machine learning (ML) are an urgent requirement of the AM industry to speed up the manufacturing process and reduce operational costs. In this context, a numerical simulation model of SLM process was developed using the ANSYS additive software. To fasten the numerical convergence, a new calibration technique was developed, which not only improves the convergence but also improves the output results. The computation efficiency is assessed and compared with the three different experimental studies. The experimental studies involve a vertical prism, a horizontal prism, and an L-shaped structure. Later another SLM numerical model developed with ANSYS workbench additive, and a datadriven framework was proposed to analyze the sensitivity and uncertainty in SLM input and output parameters. The proposed data-driven approach combines machine learning techniques with high-fidelity numerical simulations to analyze the SLM process more efficiently and this framework can be used in process optimization. Research work considers laser speed, hatch spacing, layer thickness, Young modulus, and Poisson ratio as input parameters, while the output variables are normal strains predictions in the built part. A surrogate model was constructed with a deep neural network (DNN) and polynomial chaos expansion (PCE) to create a response surface between the process output and the input variables. The analysis found that all the considered parameters were important in the process. Subsequently, the surrogate model was integrated with non-intrusive optimization algorithms such as genetic algorithms (GA), differential evolution (DE), and particle swarm optimization (PSO) to perform an inverse analysis which helped in finding the optimal parameters setting for the building. Among all the models, the PSO performed well, and the DNN model was found to be the most efficient surrogate model compared to the PCE. Another contribution from this thesis is the introduction of two data-driven, non-intrusive, reduced-order models (ROMs) named convolutional autoencoder- multilayer perceptron (CAEMLP) and a combined proper orthogonal decomposition- artificial neural network (POD -ANN). The POD-ANN uses proper orthogonal decomposition-based, reduced-order modeling which extracts the reduced order basis for the given high-fidelity input snapshot matrix. After that, an artificial neural network was constructed to form a surrogate model between the reduced order bases and the input parameters. Similarly, the CAE-MLP uses a 1D convolutional autoencoder for the reduction of a high-fidelity spatial dimension snapshot matrix constructed from high-fidelity numerical simulations. The reduced latent space is projected to the input variables using a multilayer perceptron (MLP) regression model. The efficiency and accuracy of these methods are quantified based on the thermo-mechanical analysis of an AM-built part. A good comparison between the statistical moments from the high-fidelity simulation results and the ROMs predictions confirms the ability of the proposed methods to accurately predict the SLM outputs. Additionally, the predictions are compared with the experimental results at different locations. While both models have shown a good correlation with the experimental results, the CAE-MLP performed better than the POD-ANN.

Date	2 Aug 2023
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Azzeddine Soulaïmani (Supervisor)