POD-ANN Thermal Modelling Framework for Rapid Thermal Analysis of 2.5D Chiplet Designs

While advanced packaging technologies such as 2.5D chiplet packaging offer advantages such as improved manufacturing yield, lower cost and higher transistor densities, they come with the cost of critical thermal management challenges. Thermal-aware design optimization often requires thousands of simulation runs, however, existing thermal modelling techniques face fundamental trade-offs between accuracy and computational cost that limit their utility. Conventional finite element simulations provide high-fidelity solutions but are computationally expensive. Recently, machine learning approaches have been proposed to replace conventional 3D steady-state thermal solvers. Once trained, machine learning models can provide reasonable predictions at millisecond timescales, showing great potential in accelerating design space exploration for thermal-aware chiplet placements. However, the long training time, complex architecture, and accuracy of neural networks have been the key limitations. We propose a novel technique that synergistically combines Proper Orthogonal Decomposition (POD) and Artificial Neural Networks (ANNs) to overcome these limitations. By leveraging POD to embedding the essential information from high-dimensional thermal data onto a low-dimensional subspace, an ANN can be rapidly trained to learn the low-rank representation. This compressed predictive model can then regenerate full high-resolution thermal fields orders-of-magnitude faster than rerunning simulations.