Deep material networks for lead-free solder alloys
Examensarbete för masterexamen
Applied mechanics (MPAME), MSc
Juhlin Onbeck, Gustav
In this report the Deep Material Network (DMN) method is applied to a microstructural model for SAC-305, SnAgCu with 3 weight percent silver and 0.5 weight percent copper, generated by a phase field simulation for spinodal decomposition. The DMN method homog enizes a heterogeneous microstructure for multiscale simulation purposes. Five microstruc tures from the phase-field simulations were generated, and for each micro structure a data set describing the homogenized elastic parameters was calculated. The data sets were generated using finite element simulations for each morphology. The DMN method is modified to use individual weights for each node and a bias, the modifications improved the accuracy of the network and brings the method more inline with conventional neural network structures. The new DMN method was evaluated on data sets for each of the individual microstructures and on combinations of multiple microstructure data sets. DMNs trained on data sets of multiple microstructures resulted in a much greater error than DMNs trained on individual data sets. The DMNs trained on a single microstructure reached a relative error of 2.5 - 3 %, whereas a DMN trained on a multi-microstructure set could at best reach an error of 11.5 %.
Machine learning, deep material network, multi-scale simulations, lead-free solder alloys.