Towards AI Research Agents in the Chemical Sciences

An underlying problem shared by experimental scientists is to achieve optimal behavior of their systems and arrive at new discoveries while searching over an array of controls. Accordingly, every scientific discovery may be reduced to solving a Combinatorial Optimization problem upon formulating the characteristic array of decision variables. While AI systems already excel in navigating the landscape of possible experiments, we argue herein that they will be able to drive the entire process of scientific experimental research. Especially, in response to the so-called Nobel Turing Challenge, regarding which Kitano envisioned AI Scientists, the goal of this paper is to provide a pragmatic roadmap to obtain AI Research Agents in the Chemical Sciences. We begin by reviewing the existing integration of Computational Intelligence into experimental systems, which already benefit from solving discovery/optimization problems. We mention recent discoveries in the domains of Enzymes' Design, Material Science, Quantum Mechanics, and Postharvest, in which AI systems played active roles in attaining some ground-breaking results -- thanks to being conception-free and unbiased by flawed intuition. We then devise a concrete work plan to train agents to formulate hypotheses by Deep Symbolic Reinforcement Learning, using knowledge representations based on processed scientific textbooks. We focus on the Chemical Sciences, which possess stationary Knowledge Graphs, and propose how to obtain an independent AI system at the graduate student level for ``core Chemistry''.