Machine chemists hamstrung by poor, human-produced chemistry literature

Post Reply
User avatar
funkervogt
Posts: 1165
Joined: Mon May 17, 2021 3:03 pm

Machine chemists hamstrung by poor, human-produced chemistry literature

Post by funkervogt »

Computational chemists have programmed machine learning algorithms with huge volumes of data about known chemical reactions. In other words, if you add Chemical X to Chemical Y under heat/pressure/time conditions A, you get Chemical Z. Just imagine every known chemical reaction loaded into a computer. Unfortunately, when the humans then ask the machine to PREDICT a chemical reaction that isn't in its training set, it usually fails (the human chemists confirm this by running the experiment themselves and seeing what the result is).

These continued failures aren't due to inherent limitations of computers. Rather, they attest to how incomplete and error-riddled the body of chemistry literature is. Someday, machines will be able to reliably predict chemical reactions, and they will be better than humans at chemistry. However, they need much more data and better data, which means we're going to have to re-do huge numbers of chemical reactions under closely controlled lab conditions, and to upload those results to the machines.
https://www.science.org/content/blog-po ... -chemistry
Post Reply