Zachary Ulissi: Designing New Molecules with Machine Learning

Zachary Ulissi: Designing New Molecules with Machine Learning


My group uses molecular simulations, high
through-put computing, and machine learning methods to predict the properties of small molecules
and how they’re going to interact with surfaces. We do many of these calculations in order
to get an understanding of how these properties impact various targets of interest for experimental
applications. And we collect these results in a database
and use machine learning techniques to predict properties of new molecules that we haven’t
studied yet. A recurring theme in chemical engineering
is really how you design molecules for applications. These are things like, “What sort of functional
groups are important?” and “How long your polymer chain should be.” These sorts of questions come up all the time,
and we usually address these questions sort of with intuition. A group of students work on the problem and
they sort of get an idea of what’s important, and then they make some tweaks, and they try
to optimize for a given property. But we’re starting to discover that we can
do a much better job of predicting these things up front and using information from different
sources to do this more efficiently. This is where machine learning methods have
really helped revolutionize the field of chemical engineering, have really helped get us in
this data-driven mindset. Molecular simulations are a great way to augment
the experimental data that we have. So it fits really naturally into this workflow:
we can use molecular simulations to come up with a quick estimate. If something is interesting, we do a more
detailed calculation. If it’s still interesting, we just do the
experiments, and we figure out if it’s actually worth following up on. And this sort of work flow should allow us
to really accelerate everything we do in chemical engineering. One idea that we’re working on right now
is electrochemically reducing carbon dioxide to building blocks of interest to the chemical
industry. And this idea is very powerful: we can basically
reduce waste CO2, and you can also make building-blocks that otherwise would have to come from fossil
fuels. If we’re going to solve a lot of the energy
and climate challenges that are on the horizon, we really need new chemical processes. And not just the processes, but we need the
things that go into them—like catalysts and surfactants and other ways of dispersing
nanomaterials—to really catch up and enable new ways of doing chemical transformations. Our group is working on ways to make this
process more efficient. The students at CMU have been really fantastic
for this kind of work. In chemical engineering specifically, we have
a strong background in systems engineering and mathematical techniques, a systematic
mindset. This idea of data-driven predictions is really
core to what a lot of the students in the department want to do. One of the great things of working with molecular
simulations is that as computing costs have come down and the methods have gotten better,
the time scale for getting results is quite quick, and we can very efficiently investigate
materials of interest to experimentalists. So we already have some interesting results
for inner metallics for CO2 reduction. And we’re hoping we can get these targets
out to experimentalists by the end of the year.

One Reply to “Zachary Ulissi: Designing New Molecules with Machine Learning”

Add a Comment

Your email address will not be published. Required fields are marked *