MIT Course Combines ML With Physical Systems

ML algorithms are notoriously opaque in that it can be difficult to determine why algorithms arrive at a certain prediction once models are trained. This is the reason ML models are sometimes dubbed a black box and why bias in AI is such a serious issue.

A new mechanical engineering course at MIT will teach students how to tackle this challenging problem by leveraging a combination of data science and physics-based engineering, according to a post on the MIT news site.

Machine learning with physical systems

Specifically, the “Physical Systems Modeling and Design Using Machine Learning” class (2.C01) will show how mechanical engineers can leverage their unique knowledge of physical systems to keep algorithms in check and develop more accurate predictions.

In addition, the course will give mechanical engineering students and researchers a fundamental understanding of data principles without specializing as data scientists or AI researchers. Even if they don’t dabble in data science work, this knowledge should also serve them well as they eventually manage data scientists on their teams at work.

While physical laws offer a range of ambiguities and unknowns such as temperature, humidity, and electromagnetic forces, data science can be used to predict these physical phenomena, explained MIT representative Mary Beth Gallagher.

“Having an understanding of physical systems helps ensure the resulting output of an algorithm is accurate and explainable,” she wrote.

This class will be taught concurrently with “Modeling with Machine Learning: from Algorithms to Applications” to give students the fundamentals in machine learning and domain-specific applications in mechanical engineering.

“What’s needed is a deeper combined understanding of the associated physical phenomena and the principles of data science, machine learning, in particular, to close the gap,” said Professor George Barbastathis.

“By combining data with physical principles, the new revolution in physics-based engineering is relatively immune to the ‘black box’ problem facing other types of machine learning.”

Students are expected to develop a final project by identifying a real-world problem that requires data science to address the ambiguity inherent in physical systems. This entails sourcing relevant data, selecting the appropriate ML method, and implementing the chosen solution.

The 2.C01 course had its first run in spring last year.

Image credit: iStockphoto/Ilya Lukichev