If I was in the position I was 35 years ago about to start a PhD in Chemistry (I decided just to stick with programming) this is exactly what I would focus on. It's not easy (though a lot easier today than back then) and might take a long time to make it comprehensive, but eventually you are going to have computers deciding on what you should focus your human-powered chemistry on. That to me would be a fascinating thing to work on.
I am there, I use Molecular Dynamic simulations to study proteins, and I have to agree with you.
Chemistry is an umbrella term---It is the summation of all the little tidbits of information we have accumulated about molecular systems. As such, the field has become unwieldy. The future lies in using computers to handle this vast amount of knowledge.
How will this work? I don't know. But I can tell you it will not be simulations, which is what everyone in academia gravitates toward. Simulations involve lots of sampling, making the approach computational intractable. So I am naively optimistic about using machine learning. For example, I can imagine using recurrent neural networks to handle carbon chains and newly invented attention models for things like protein folding.
P.S. Now I just need to find a way to fund this research :-(
One interesting aspect of statistical machine learning is the cornucopia of techniques related to the "exploration versus exploitation" dilemma in exploring a giant search space. Cf. "bandit learning"
I think it will take a novel combination of many different aspects of data mining, machine learning, and network science to make a major breakthrough, but it also definitely sounds like a really fun problem to work on.
4 comments
[ 3.1 ms ] story [ 16.2 ms ] threadChemistry is an umbrella term---It is the summation of all the little tidbits of information we have accumulated about molecular systems. As such, the field has become unwieldy. The future lies in using computers to handle this vast amount of knowledge.
How will this work? I don't know. But I can tell you it will not be simulations, which is what everyone in academia gravitates toward. Simulations involve lots of sampling, making the approach computational intractable. So I am naively optimistic about using machine learning. For example, I can imagine using recurrent neural networks to handle carbon chains and newly invented attention models for things like protein folding.
P.S. Now I just need to find a way to fund this research :-(
I think it will take a novel combination of many different aspects of data mining, machine learning, and network science to make a major breakthrough, but it also definitely sounds like a really fun problem to work on.