Upscaling physics simulations at system-level with AI and finite elements

Physics simulations are ubiquitous in chip design: electromagnetic interactions, thermal transport, mechanical deformation, fluid cooling… With new design paradigms (novel 3D packaging and integration mechanism to stack logic, memory and I/O), physics simulations are required at the system-level to push optimization further across the stack. But as individual components shrink in size, this results in an impractical scale range for high-fidelity simulation tools. Machine Learning (ML) algorithms have shown great potential to provide computationally efficient surrogate models for physical processes but suffer from the limited availability of reliable training data when it comes to large-scale problems.

We propose to derive a novel hybrid paradigm that combines machine learning with traditional computational physics methods. It consists in learning a functional representation of subdomains (typically one component/device) and apply the finite element method to upscale at the system level (thousands to millions  of components). At the component scale, the training process is fast and data are abundant. At the system scale, the upscaling is reliable as physical laws are enforced.

The goal of the PhD project is to develop such a framework to address thermal and mechanical simulations of 3D semiconductor chip packaging. What are the best machine learning algorithms for modelling component response? To what extent are the component models reusable for different designs and physical phenomena? What are the trade-offs between accuracy and computational efficiency in this new paradigm? These are the scientific questions this PhD will seek to answer.