What’s Heating Up Your iPhone? Understanding the Thermal Physics of Microchips With AI

In the era of hybrid cloud, AI, and the Internet of Things, the demand for faster semiconductor microchips continues to rise. To make microchips faster, it is necessary that we include billions or even trillions of nanometer-scale transistors onto a fingernail sized chip. Transistors are the fundamental building blocks of microchips: more the number of transistors, faster the microchip. The state-of-the-art transistors include layers of semiconductors, dielectric materials and metals and the thickness of the layers within the structure can be in the 1 nm to 5 nm range. The smaller and smaller dimensions of the transistors cause the heat to get trapped inside them, which leads to overheating, and in turn, defect generation, performance loss, and outright failure of the microchips. The existing thermal models often approximate the thermal properties of these extremely complex structures using bulk material properties. However, the thermal physics of highly confined nanoscale materials can be drastically different from their bulk counterparts. The existing models do not consider such deviations and as a result, cannot suggest what needs to be done to keep the transistors from overheating. On the other hand, atomistic simulations have shown remarkable accuracy in predicting heat transport properties of simple nanoscale materials. However, the simulations usually require days to weeks for even simple structures and they cannot be directly used to model these transistors.

 

In this talk, I will discuss how atomistic modeling combined with machine learning methods allows us to overcome this challenge and predict thermal properties of sub-10-nm transistors. We analyze the heat transport properties of transistors, using Graphics Processing Units (GPUs) Molecular Dynamics, a highly efficient general-purpose molecular dynamics package fully implemented on GPUs and machine-learned potentials called neuroevolution potentials (NEPs). Our AI-accelerated modeling provides physical insights to rapidly achieve thermally optimized transistor designs and in turn, fabricate faster and energy-efficient semiconductor microchips. We gratefully acknowledge funding from the Defense Advanced Research Projects Agency, Microsystems Technology Office, for the work.