Catalyst control is critical to carbon nanotube (CNT) growth and scaling their production. In supported catalyst CNT growth, the reduction of an oxidized metal catalyst enables growth, but its reduction also initiates catalyst deactivation via Ostwald ripening. Here, we conducted autonomous experiments guided by a hypothesis-driven machine learning planner based on a novel jump regression algorithm. This planning algorithm iteratively models the experimental response surface to identify discontinuities, such as those created by a material phase change, and targets further experiments to improve the fit and reduce uncertainty in its model. This approach led us to identify conditions that resulted in the greatest CNT yields as a function of the driving forces of catalyst reduction in a fraction of the time and cost of conventional experimental approaches. By varying temperature and the reducing potential of the growth atmosphere, we identified discontinuous jumps in CNT growth for two thicknesses of an iron catalyst, resulting in largest observed yields in narrow and distinct regions of thermodynamic space where we believe the reduced catalyst is in equilibrium with its oxide. At these jumps, we also observed the longest growth lifetimes and a greater degree of diameter control. We believe that conducting CNT growth at these conditions optimizes catalyst activity by inhibiting Ostwald ripening-induced deactivation, thereby keeping catalyst nanoparticles smaller and more numerous. This work establishes a thermodynamic framework for a generalized understanding of metal catalysts in CNT growth, and demonstrates the capability of iterative, hypothesis-driven autonomous experimentation to greatly accelerate materials science.
