PI-ONet: A Physics-Informed Operator Network for Efficient Thermal Analysis of Multilayer Chiplets

With the continued advancement of integrated circuit technology, chiplet design is becoming increasingly important within the semiconductor industry. Concurrently, as the number of small chiplets and their integration densities increase, the complexity of thermal design and management escalates. Traditional thermal simulation methods, such as the finite element method (FEM), struggle to meet the demands for rapid design and optimization due to their high computational complexity and time consumption. This study introduces a novel operator neural network model, the Physics-Informed Operator Network (PI-ONet), which is designed to overcome these challenges. By incorporating fundamental physics into the training process, PI-ONet not only enhances the accuracy of predictions but also bolsters reliability. PI-ONet comprises two components: the chiplets temperature prediction model, which learns the mapping from power distribution to temperature distribution, and the thermal conductivity equivalent model, which simulates the thermal conductivity of the bump layer and the insertion layer. Experimental results demonstrate that PI-ONet substantially increases the efficiency of thermal simulations, achieving an acceleration of up to 35,000 times compared to the traditional FEM tool Icepak. This remarkable enhancement in efficiency highlights the potential of PI-ONet in the thermal design and management of complex structures and multi-chiplet systems, suggesting promising application prospects in advancing efficient chiplet system design.