Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Application of machine learning in multiscale thermal modelling of laser powder-bed fusion processes
Finite element (FE) thermal simulation of additive manufacturing processes is computationally very expensive due to the need for employment of fine time and space discretization levels. This student project aims to employ neural networks as a general function approximator to develop a metamodel that would generate the outcome of FE simulations based on given inputs parameters.
Thermal simulation of powder bed fusion additive manufacturing (PBFAM e.g. SLM & EBM) is a challenging task. It involves calculation of a highly localized transient temperature field generated by a moving laser or electron beam with a typical radius of 30-100 μm and a velocity of 100-1000 mm/s. Reliable finite element models of these processes require fine space and time discretization in the order of micrometres and microseconds, respectively. On the other hand, PBFAM builds are typically in the range of centimeters and their manufacturing takes hours. This discrepancy between the scale of conventional PBFAM simulations and actual parts makes component-scale simulations inaccessible.
A multiscale solution is being developed as a part of a PhD project to decrease the computational cost of the thermal simulations of PBFAM and enable part-scale modelling. The approach combines simulation results from a global coarse-mesh and thousands of local fine-mesh FE models to obtain an accurate description of the temperature distribution during PBFAM of a part. The main computational burden of this approach is the FE-solution of thousands of local fine-mesh models which are identical in geometry but differ in their initial and boundary conditions. The present student project aims to examine the effectiveness of neural networks in developing a data-driven surrogate model for the well-defined local FE simulations. The metamodel would be trained based on a varied set of local models to predict the outcome of any given set of input parameters that can occur in the multiscale FE modelling approach. Thus, the overall computational costs would be substantially reduced by evaluating only some of the thousands of local models.
This project is the continuation of a series of master theses revolving around the same topic which would serve as the starting point.
This project is suitable for **ETH** MSc thesis and for students who are passionate about applying machine learning concepts in the field of mechanical engineering and are interested in programming and finite element analysis.
Thermal simulation of powder bed fusion additive manufacturing (PBFAM e.g. SLM & EBM) is a challenging task. It involves calculation of a highly localized transient temperature field generated by a moving laser or electron beam with a typical radius of 30-100 μm and a velocity of 100-1000 mm/s. Reliable finite element models of these processes require fine space and time discretization in the order of micrometres and microseconds, respectively. On the other hand, PBFAM builds are typically in the range of centimeters and their manufacturing takes hours. This discrepancy between the scale of conventional PBFAM simulations and actual parts makes component-scale simulations inaccessible.
A multiscale solution is being developed as a part of a PhD project to decrease the computational cost of the thermal simulations of PBFAM and enable part-scale modelling. The approach combines simulation results from a global coarse-mesh and thousands of local fine-mesh FE models to obtain an accurate description of the temperature distribution during PBFAM of a part. The main computational burden of this approach is the FE-solution of thousands of local fine-mesh models which are identical in geometry but differ in their initial and boundary conditions. The present student project aims to examine the effectiveness of neural networks in developing a data-driven surrogate model for the well-defined local FE simulations. The metamodel would be trained based on a varied set of local models to predict the outcome of any given set of input parameters that can occur in the multiscale FE modelling approach. Thus, the overall computational costs would be substantially reduced by evaluating only some of the thousands of local models.
This project is the continuation of a series of master theses revolving around the same topic which would serve as the starting point.
This project is suitable for **ETH** MSc thesis and for students who are passionate about applying machine learning concepts in the field of mechanical engineering and are interested in programming and finite element analysis.
The objective of the project is to train and validate a data-driven alternative for thermal simulation of the selective laser melting process.
The objective of the project is to train and validate a data-driven alternative for thermal simulation of the selective laser melting process.