Data driven optimization of structural and dynamical properties of parts manufactured via selective laser melting process using machine learning and finite element analysis methods

Selective Laser Melting (SLM) is a member of Powder Bed Fusion (PBF) process which is one of the most popular branches in metal Additive Manufacturing (AM) production technique. SLM have advantageous characteristics in terms of design flexibility, wide range in selection of material, relatively small amount of feedstock wastage, minimum lead time and assembly needs etc. This method have different processing parameters that plays significant role in the process quality such that Laser Power, Scanning Speed, Hatch Distance, Laser Beam Diameter, Layer Thickness can be adjusted in the process to get better structural performance from parts. SLM method is a thermo-mechanical process so that it can be modelled with numerical applications such as Finite Element Analysis (FEA), Discrete Element Method (DEM) etc. to improve the structural dynamics of produced parts and to understand the process better. In addition, predictive models can be developed based on data sets with Artificial Intelligence (AI) methods such as Machine Learning (ML). ML is an efficient way to create correlations between inputs and output(s) and it can be developed based on existing dataset. Thus, ML algorithms and FEA applications can be used as joint methods on optimizing processing paramaters and process modelling of SLM parts. By its nature, SLM is a nonlinear process. Therefore, the correleation between multiple inputs and multiple outputs for the multi-objective and constrained optimization method is subject of this thesis. The main purpose of the study is to modify structural dynamics of parts such as relative density, elasticity, surface roughness, natural frequency, mode shapes by using ML - FEA and create prediction - optimization model regardless of selected material. In the scope of the study, mechanical properties of SLM parts were predicted by different ML algorithms such as Artificial Neural Network, Support Vector Machines, Gaussian Process Regression. In this manner, together with composing deep literature survey, an appopriate Design of Experiments strategy implemented. Generated models were combined with FEA which are subsequently used in multi-objective constrained optimization algorithms e.g. Sequential Quadratic Programming (SQP), Hill Climbing method. The optimization constraints are formed according to the limits of the process parameters. Consequently, this study proofed that multiple objective optimization methods can be successfully tailored with ML and FEA to produce high structural performance parts. It is also validated that natural frequencies are function of structural properties but mode shapes were not affected by changing material constants. Furthermore, results verified that the optimization process can be applicable in different novel alloys such as AlSi10Mg, SS 316L, Inconel 718 which validates the proposed model works as material independent. Lastly, ML, FEA and SQP methods have reasonably compatible performance to each other in terms of optimization process outcomes.