Introduction

You may have encountered or heard about some puzzling limitations of AI despite its impressive capabilities, especially when applying AI in specific domains such as scientific discovery and manufacturing: (i) struggling to exploit fine-grained domain knowledge; (ii) lacking the controllability on high-stakes tasks; (iii) remaining gaps between established workflows. These challenges reflect a fundamental trade-off in AI: breadth and depth cannot be simultaneously maximized—there is no free lunch.

To address these limitations, researchers have developed Domain-Specific Representations (DSRs)—formal artifacts that help AI systems better incorporate domain knowledge. However, designing effective DSRs requires in-depth collaboration between domain experts and computer scientists, making the process labor-intensive, case-specific, and thereby costly.

This project aims to democratize DSR design by developing intelligent algorithms that can automate this process. Specifically, we will investigate: (i) algorithms for automated DSR design—concretely, our algorithms will take diverse sources of domain knowledge (including standard operating procedures and experts' tacit knowledge) and output DSRs implemented as domain-specific programming languages; (ii) evaluation criteria for both the design algorithms and the resulting DSRs across cutting-edge facilities in scientific discovery and manufacturing; and (iii) integration strategies for deploying AI systems with these DSRs in applications, including self-driving laboratories that autonomously conduct experiments, dark factories that operate without human presence, and coordinated teams of heterogeneous unmanned devices.

Together, we'll work to make DSRs accessible to a broader spectrum of domains, pioneering new approaches at the intersection of AI and specialized knowledge.