Scalability of automated solver translation

Determine whether software tools and agentic pipelines for automatically translating numerical solvers to accelerators can scale to complete simulation pipelines.

Background

The thesis discusses automated code-generation and translation systems as a possible alternative to learning neural surrogates: existing CPU solvers could potentially be ported directly to GPUs while preserving their numerical behavior and adding automatic differentiation and vectorization.

Although early work has automated portions of this process, the thesis identifies the scalability of such tools to full simulation workflows as unresolved. The problem concerns whether automated translation can handle complete, heterogeneous numerical pipelines rather than isolated kernels.