Abstract

Additive manufacturing (AM) processes are among the manufacturing methods implemented in various industries. Direct metal deposition (DMD) is part of AM processes that uses the laser heat source to deposit the metallic material in the form of powder or wire onto a substrate and build a component in a layer-by-layer scheme. The DMD process is known to be cost-effective and easily adaptable for building complex structures. During a DMD process, material experiences several heating and cooling cycles which lead to the formation of residual stresses and distortions of the fabricated part.
There are several experimental-based methods and techniques for measuring the residual stresses of metallic components. However, the application of these methods can damage the fabricated parts or may require considerable time and tooling expenses for the experiment. Alternative solutions such as finite element (FE) analysis were developed to predict the residual stresses without damaging the part. The application of the FE in assessing the residual stress distribution is time-efficient and cost-effective. The FE analysis of DMD process includes thermal and mechanical analyses; the temperature history of the elements is obtained by performing a pure heat transfer analysis, then it is applied to the mechanical model to calculate the structural response of the part. One of the shortcomings of the FE analysis of DMD process corresponds to the high computational time of the mechanical analysis. Therefore, several techniques and approaches were developed in the literature to address this issue and improve the computational efficiency of the FE method.
Throughout this thesis, a novel approach of integrating the FE analysis with artificial neural networks (ANNs) is presented as an efficient method for improving the computational time of predicting the residual stresses in DMD fabricated parts. ANNs are part of machine learning (ML) algorithms that tries to determine the logical relationship between the given inputs and the associated output(s). A feed-forward ANN with gradient descent backpropagation developed in Keras was implemented. The ANN is trained by feeding the dataset into the network and minimizing the error function.
In the present study, several structures made from AISI 304L with 12-layers and 18-layers deposition were considered. and a detailed thermomechanical FE analysis was performed on them. Temperature history of the elements along with their dimensional features of 12-layers structures were extracted as the inputs and the corresponding residual stress components were recorded as the outputs to train the ANN. On the other hand, the temperature history of the elements and their geometrical features extracted from 18-layers structures were fed into the trained ANN for making predictions. The results of the integrated ANN-FE are compared with the results of the residual stresses of 18-layers obtained from the detailed thermomechanical analysis. The prediction errors were calculated and shown in the form of 3D contours and scattered errors. Moreover, the histogram analysis was performed for each 18-layers structures to better present the fraction of the elements with the associated error ranges. Finally, the computational times are recorded and compared with the results of the detailed FE analysis to evaluate the efficiency and performance of the proposed novel ANN-FE method.
The results showed that for almost all of the structures and all the stress components, the predicted pattern and magnitude of the residual stress were consistent with the detailed FE analysis. For some of the structures, very high errors were observed which were associated with the low-stress state zones in which the actual stress magnitude was low and the high errors pose no critical condition. Although there are some predictions showing higher errors in some regions, the majority of the elements in the structures showed prediction errors of less than 15% supported by the histogram analysis. Significant improvement in the computational time of the 18-layers structures was also achieved (6 times as an average). The computational time of predicting the residual stresses in the DMD parts was improved substantially with low loss in the accuracy of the predicted results. Therefore, the proposed method can be implemented for investigating the effects of the hyperparameters on the residual stresses in DMD process.