Learning based mesh generation for thermal simulation in handheld devices with variable power consumption

This work utilizes machine learning techniques to achieve learning based mesh generation for the thermal simulator. In the learning stage, first, we generate many power value sets of application processors and apply the superposition technique to obtain the temperature raw data of material blocks. We, then, use these thermal data and several defined thermal gradient ranges to obtain different appropriate mesh-size sets of material blocks for each power value set. We construct the regressions that are capable of predicting targets as a supervised learning task, and the architecture of regression model includes seven hidden layers. After the learning stage, we integrate the learning based mesh decision engine into a developed system-level thermal simulator, which is built by the finite difference method. The experimental results show that the temperature profiles of the handheld device calculated by the developed thermal simulator with learning based mesh generation fits the results obtained by a commercial tool pretty well with less runtime usages. Moreover, the learning based mesh decision engine is much more efficient than the artificial mesh decision.