NSF Award
2326174
Computer programming is essential for modern science. Scientists write programs to control instruments, run simulations, and analyze data. However, the programming tools and techniques that scientists use often lag behind those in the software engineering industry. This lag makes scientific discovery slower, more costly, and can lead to unreproducible results. In the past two years, artificial intelligence (AI) tools, such as ChatGPT, have revolutionized the software industry. They have been shown to make software engineers significantly more productive, but have not had the same impact on the sciences. The goal of this research project is to develop and test AI programming tools that work for scientists. The research team is developing AI models and tools that support the programming languages that scientists use. They are developing benchmarks to evaluate the effectiveness of AI tools for programming tasks that are unique to the sciences. They are investigating how AI programming tools can help college students study science more effectively. By harnessing AI to make programming easier for scientists, the project is helping to accelerate scientific discovery, lower its cost, and allow more people to participate in scientific work.<br/><br/>The project is developing large language models of code and associated tools to support scientists. To understand scientists' needs, the team is running qualitative and quantitative studies of how scientists write programs. Based on these findings, they are developing deep neural network models for programming languages that are frequently used in the sciences, such as MATLAB and R, but are less commonly used in the software engineering industry. These models are particularly helpful for scientists who are not expert programmers; they can turn descriptions into computer programs, and also generate explanations of existing programs. The team is developing models that support the programming paradigms that scientists use, including computational notebooks and programs whose structure is determined by data formats. The team is developing code generation models that can be deployed on private, "air gapped" networks, making them suitable for scientists working in sensitive fields, including energy and defense.<br/><br/>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.
