Teaching Machine Learning

Machine learning is often thought of as the silver bullet for solving puzzles in science. With enough data in any given subject, a model with some impressive predictive properties can theoretically be created to produce an abundance of helpful information. But sometimes, machine learning isn’t perfect. In their recent paper, Guannan Liu and his colleagues at the Schroers Lab at Yale explore the limitations of machine learning models and how we can incorporate human learning into new models to strengthen their predictive power. 

Liu, a Ph.D. candidate in Mechanical Engineering and Materials Science, has spent most of his time at Yale studying machine learning and its capability to solve complex materials science problems. The study of glass-forming ability is a canonical example of one of these problems. It is quantified by the minimum cooling rate required to prevent the formation of undesired crystalline structures, resulting in a glass with an amorphous atomic structure. Studying glass-forming ability by performing hands-on experiments in the lab can be tedious, and this is where machine learning comes in to potentially accelerate the process. “Simply put, machine learning is trying to make inferences from data, and maybe predict something from [that] data,” Liu said. 

But there’s a catch—previous machine learning models designed to predict the glass-forming ability of metallic glasses have fallen short of providing useful insights on the subject. Metallic glasses are alloys, which are made by combining two or more elements. “You have to have meaningful features that describe the particular alloy after the mixing of elements,” Liu said. Liu contextualized this idea by offering an example. “For atomic size, it’s not the average [element size] that matters but the difference in size,” Liu said. “This allows space to be filled as much as possible, favorable for glass formation.” Models that lack such information and instead arbitrarily use statistical functions to construct features do not truly capture the essence of the alloy’s glass-forming ability. 

The other issue with the previous machine learning model for the glass-forming ability of alloys was its limited capacity to make new predictions.  “[In] the previous model, usually the task was interpolation–the model was only predicting things that were similar to the dataset,” Liu said. In other words, the model could not make any predictions for new and unfamiliar data—it could only work inside the bounds of the dataset. 

To rectify these errors, Liu and his team worked on a new machine learning model: one that incorporated scientific insights from human learning. “Our model used extrapolation, [or] prediction into unknown space,” Liu said. This way, they were able to align their machine learning model with the reality that they have observed in the lab. Take, for instance, the property of atomic size again. Larger differences in size result in better glass-forming ability because the atoms are able to pack in more tightly. It is properties such as these that Liu and his team were better able to account for in their model, and their approach worked. “We found that our model was actually very successful in predicting glass-forming ability,” Liu said. 

Unlike its predecessors, this model was much better at extrapolation. “Our model can predict alloys that are more distinct from the training set,” Liu said. “We concluded that physical insights are really needed [to develop effective machine learning models].”

This project is far from the end of Liu’s work with machine learning for materials science. He has three main goals moving forward. “The first is to use machine learning to test [our] current understanding of complex material science problems,” Liu said. It can be hard to quantify the efficacy of foundational rules within material science, so Liu plans to use machine learning to evaluate these guiding principles. Secondly, Liu strives to combine machine learning and high-throughput fabrication methods to discover new metallic glasses. 

Finally, Liu will investigate the contexts in which machine learning can be helpful. His goal is to determine what kinds of problems it can help solve and what circumstances limit its utility. “[We want to] have a viewpoint that can be meaningful for the community as to what machine learning is useful for [versus] situations where machine learning would be hard to use,” Liu said. This research will ensure that machine learning models are taking all possible human insights into account while making inferences from data. Clearly, machine learning can be an extremely valuable tool if it is wielded skillfully.