I wonder if another factor to this is the lack of published/available "negative" data (eg. reacting A + B _does not_ make AB even though the very similar A + [C-F] reactions do successfully create A[C-F]).
A lot of chemical reactions are already susceptible to small changes in reaction profile and while there _are_ human heuristics for dealing with these I'm not sure that just learning from successful reactions would allow you to derive these.
TFA does mention this but it's already a problem that humans face with duplicated effort of repeating reactions that are known to fail by someone else. _No one_ is publishing this stuff and starting now probably won't fill the data void.
A lot of chemical reactions are already susceptible to small changes in reaction profile and while there _are_ human heuristics for dealing with these I'm not sure that just learning from successful reactions would allow you to derive these.
TFA does mention this but it's already a problem that humans face with duplicated effort of repeating reactions that are known to fail by someone else. _No one_ is publishing this stuff and starting now probably won't fill the data void.