Lookup qsar and qspr. They’ve been doing this for decades with various statistical methods to hunt for new perfume (or drugs or whatever). While the statistical tools look (far?) better now it’s still a data driven problem: how much and how good. Also how pertinent are the descriptors that are used and whether any prediction can be made successfully outside of the training set.
They also hope their model(s) of the olfactory system are good enough as it’s basically a drug design problem to a receptor system.