%0 Generic %1 hornischer2021reduction %A Hornischer, Niklas %D 2021 %E Kässinger, Johannes %E Rosin, David %K PN7-1 PerSiVal simtech %T Mesh Size Reduction and Interpolation of a Biomechanical Simulation for Neural Networks on Mobile Devices %X In simulation science, the application and visualization of complex models and calculations is often limited by resource-constrained hardware. To overcome this issue of restricted access to simulation results and allow for pervasive simulations, this limitation has to be overcome. In the context of the Digital Human Model, a complex biomechanical simulation of a muscle should be visualized in real-time on a mobile device. In that way, the usage of a distributed system and neural networks to predict the simulation results points out several difficulties regarding computational osts and network latency. This project work proposes a interpolation model to construct the surface of the M. biceps brachii out of a few given positions on the surface. We use dense neural networks to predict the differences of the actual surface to a reference, based on the differences between the few given positions on the surface and the reference. To find the best subset of reduced positions genetic algorithms and mesh simplification methods are used. Our interpolation model is able to interpolate from 30 nodal points to 2809 nodal points by using only one dense layer with a linear activation function. Experiments reveal, that this specific interpolation model achieves a mean absolute error of 0.046 mm and reaches 30.46 FPS in computational time on a local hardware. Overall, the interpolation model achieves real-time performance while ensuring high accuracy of the interpolation results and decreased network latency due to extensive reduction of the transmitted data.

@misc{hornischer2021reduction, abstract = {In simulation science, the application and visualization of complex models and calculations is often limited by resource-constrained hardware. To overcome this issue of restricted access to simulation results and allow for pervasive simulations, this limitation has to be overcome. In the context of the Digital Human Model, a complex biomechanical simulation of a muscle should be visualized in real-time on a mobile device. In that way, the usage of a distributed system and neural networks to predict the simulation results points out several difficulties regarding computational osts and network latency. This project work proposes a interpolation model to construct the surface of the M. biceps brachii out of a few given positions on the surface. We use dense neural networks to predict the differences of the actual surface to a reference, based on the differences between the few given positions on the surface and the reference. To find the best subset of reduced positions genetic algorithms and mesh simplification methods are used. Our interpolation model is able to interpolate from 30 nodal points to 2809 nodal points by using only one dense layer with a linear activation function. Experiments reveal, that this specific interpolation model achieves a mean absolute error of 0.046 mm and reaches 30.46 FPS in computational time on a local hardware. Overall, the interpolation model achieves real-time performance while ensuring high accuracy of the interpolation results and decreased network latency due to extensive reduction of the transmitted data.}, added-at = {2021-12-13T13:36:53.000+0100}, author = {Hornischer, Niklas}, biburl = {https://puma.ub.uni-stuttgart.de/bibtex/2e0602b4b97c014ecb709281acb13558f/davidrosin}, description = {This is a student-thesis submitted within the bachelor program "Simulation Technology".}, editor = {Kässinger, Johannes and Rosin, David}, interhash = {55b95422b39d9aa4da2f653381e2bce2}, intrahash = {e0602b4b97c014ecb709281acb13558f}, keywords = {PN7-1 PerSiVal simtech}, timestamp = {2021-12-13T13:10:07.000+0100}, title = {Mesh Size Reduction and Interpolation of a Biomechanical Simulation for Neural Networks on Mobile Devices}, year = 2021 }