Chemical Engineering Professor Receives Prestigious CAREER Award

Engineers must make predictions to design something new. A chemical engineer planning a storage tank needs to know how a gas or liquid will behave as temperature and pressure change. A biomedical engineer formulating a new pharmaceutical may have to predict how much of a drug compound will dissolve in water or other liquids and how temperature affects that process. 

Engineers use equations of state — specific mathematical representations of physical and chemical properties — to make those predictions. But an equation of state is only as accurate as the data on which it's based. Collecting more data can be difficult and expensive. And an equation of state does not tell an engineer how certain its predictions are. 

Jacob Monroe, an assistant professor of chemical engineering, has received a five-year National Science Foundation CAREER Award worth $563,155 to use machine learning to develop a model that can make more accurate predictions of thermodynamic properties while also providing estimates of uncertainty. While Monroe is a chemical engineer, the model could improve the work of engineers in many fields. 

The prestigious CAREER Awards support early-career faculty with the "potential to serve as academic role models in research and education," according to the NSF. 

"Jacob was an incredible hire for us back in 2023," said Bob Beitle, interim head of the Ralph E. Martin Department of Chemical Engineering. "His CAREER proposal successfully integrated all components of faculty career development; it is no wonder it was awarded." 

Learning About Machine Learning 

Monroe was a University of Virginia undergraduate working in a biomedical lab when he saw the possibilities of using computers to speed the research process. The lab was screening potential drugs. What if the computer could narrow down target molecules, making the lab work more efficient? 

He discovered that this was already an established area of research. 

"It felt very personal, like I came up with this idea. And then someone said, 'oh yeah, people do this.' And I never turned back," he said. 

In his research, Monroe looks at individual atoms to predict a system's thermodynamic properties, such as temperature, pressure and density. 

During his postdoctoral fellowship, his adviser encouraged him to investigate whether machine learning could aid his work. 

"I was kind of skeptical of machine learning. But the more I read on it, I saw it was tightly tied to what we call statistical mechanics," he said. 

Imagine you want to know the thermodynamic properties of a glass of water. It's not possible to simulate every atom in the glass. Using statistical mechanics, Monroe can study a small number of molecules and predict the thermodynamic properties of the entire glass of water. Machine learning can follow a similar approach, using a smaller set of known data to predict other results. 

"I try to go back to the theory and come up with new methods that improve simulations and our ability to make physical predictions by incorporating machine learning techniques," Monroe said. 

A Tool for Working Engineers 

The equation of state Monroe is creating can make better predictions of thermodynamic properties when data is limited. The method can also point to what additional data would most improve the predictions. That allows engineers to focus their resources on collecting the data that would have the greatest impact. 

"Before running actual experiments, let's try running simulations," Monroe said. "And we can run a lot of simulations on supercomputers. Instead of running, say, 100 experiments, you might be able to run 10,000 simulations and then only need three experiments." 

The model will also tell engineers how certain its predictions are. 

"If you have an important decision you need to make and you don't have a model that gives you high certainty, you can go back and collect more data," he said. 

A degree of uncertainty might be acceptable in some situations. For others, it could be dangerous. 

The software that engineers use today to design products and systems does not provide an estimate of uncertainty. One day, Monroe believes, his research could be incorporated into those programs and become an integral part of how engineers carry out their daily work. 

"We have a long way to go. We have to do the work, do the research. We have to talk to people who create these software programs and convince them this is a useful thing," Monroe said. "But I see this becoming extremely beneficial."