Collaborative Research: Research Infrastructure: CIRC: ENS: Bridging Science and AI Through Differentiable Program Transformations

Automatic Differentiation has become an important enabling technology for scientific computing and machine learning. Simply put, machine learning is a parameter fitting problem, and the computation of derivatives enables the fitting of parameters. Historically, programmers were required to spend significant time and effort developing these derivative codes by hand, making the use of machine learning and simulation on existing applications a tedious and difficult task.

In recent years tools have been developed to generate code to compute these properties. However, they have been limited to specialized domains and specific programming languages. In contrast, the Enzyme project aims to generate derivatives of arbitrary programs in any LLVM-based language (e.g.

C/C++, Fortran, Julia, Rust, Swift, Python, MLIR, etc), without restriction on scientific domain. The project's impacts are that scientists and engineers in all fields will be able to apply modern algorithms like neural networks to their domains without extensive rewriting of their entire application.

This project will both develop the existing research prototype Enzyme implementation into a production-quality ecosystem, and establish an open-source community that will allow Enzyme to be maintained by the open-source community. This involves extensive user documentation and examples, documentation for new and existing Enzyme developers, integration into vendor compilers, organizing an annual Enzyme conference and weekly developer meetings, providing tutorials for Enzyme at relevant meetings, creating an Enzyme advisory board, and coordinating community satisfactions as well as development priorities.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.