Static chip thermal analysis provides detailed and accurate thermal profile on chip. The chip power map, commonly modeled as rectangular regions of distinct heat sources, significantly impacts the chip thermal profile. Since the heat sources result from numerous cells in functional blocks, the design space of chip power map is prohibitively enormous. Numerical simulations can be reliable for solving complex power maps; however, it could be very time-consuming when simulating a large SoC and/or 3DIC designs. Thus, there is an urgent need for speeding up the static chip thermal analysis to tackle various power maps. In this paper, we propose an approach of integrating our developed machine learning thermal solver [1] and decay curve characterization for solving static chip thermal with diverse power maps. The machine learning thermal solver would first solve the power maps on a coarse level (e.g., 200 um). The thermal results are further enhanced using the decay curve algorithm which would fine tune the solution locally provided by the machine learning thermal solver and calculate the local temperature variations at a finer level (e.g., 10 um). The deep learning models are trained on augmented artificial power maps and tested on realistic chip power maps. Experimental results validate the effectiveness of the proposed approach of offering fast and accurate chip thermal profile.

In this paper we show that our Machine Learning (ML) approach, CoMLSim (Composable Machine Learning Simulator), can simulate PDEs on highly-resolved grids with higher accuracy and generalization to out-of-distribution source terms and geometries than traditional ML baselines. Our unique approach combines key principles of traditional PDE solvers with local-learning and low-dimensional manifold techniques to iteratively simulate PDEs on large computational domains. The proposed approach is validated on more than 5 steady-state PDEs across different PDE conditions on highly-resolved grids and comparisons are made with the commercial solver, Ansys Fluent as well as 4 other state-of-the-art ML methods. The numerical experiments show that our approach outperforms ML baselines in terms of 1) accuracy across quantitative metrics and 2) generalization to out-of-distribution conditions as well as domain sizes. Additionally, we provide results for a large number of ablations experiments conducted to highlight components of our approach that strongly influence the results. We conclude that our local-learning and iterative-inferencing approach reduces the challenge of generalization that most ML models face.