Authors: Andres Munoz-Jaramillo (Southwest Research Institute)

The application of Artificial Intelligence (AI) to solve scientific objectives contains an intrinsic tension between versatility and interpretability.  On the one hand, the main purpose of science is human understanding, on the other hand, the power of AI resides in its ability to approximate mathematical relationships using multiple layers of learnable abstraction (a process that is incomprehensible to humans).   This creates a tension that many scientists find distasteful and difficult to navigate.   However, our ability to use mathematics to describe reality is the main the pillar upon which progress has been built and AI (as a mathematical construct that allows us to use data to approximate mathematical relationships to an arbitrary degree of precision) has tremendous potential to revolutionize science.

Here we discuss the different elements that enable an AI system to approximate mathematical relationships, how this can leveraged to maximize the benefit that multiple observations can bring to a problem, and how the most interesting applications of AI involve careful design and application of the black box to specific parts of the system, while keeping others to well understood physics engines.