<p></p><p></p><p>We present a robotic chemical discovery system capable of
learning the generalized notion of reactivity using a neural network model that
can abstract the reactivity from the identity of the reagents. The system is controlled
using an algorithm that works in conjunction with this learned knowledge, the
robot was able to autonomously explore a large number of potential reactions
and assess the reactivity of mixtures, including unknown datasets, regardless
the identity of the starting materials. The system identified a range of
chemical reactions and products, some of which were well-known, some new but
predictable from known pathways, but also some unpredictable reactions that
yielded new molecules. The search was done within a budget of 15 inputs
combined in 1018 reactions, which allowed us not only to discover a new
photochemical reaction, but also a new reactivity mode for a well-known reagent
(<i>p</i>-toluenesulfonylmethyl isocyanide, TosMIC). This involved the reaction
of six equivalents of TosMIC in a ‘multi-step, single-substrate’ cascade
reaction yielding a trimeric product in high yield (47% unoptimized) with
formation of five new C-C bonds involving <i>sp</i>-<i>sp<sup>2</sup></i> and <i>sp</i>-<i>sp<sup>3</sup></i>
carbon centres. Analysis reveals that this transformation is intrinsically unpredictable,
demonstrating the possibility of reactivity-first robotic discovery of unknown
reaction methodologies without requiring human input.</p><br><p></p><p></p>