Physics-based engineering and the “black box” problem of machine learning

Through

In MIT 2.C161, George BarbastatThis shows how mechanical engineers can use their knowledge of physical systems to control algorithms and develop more accurate predictions.

Machine learning algorithms are often referred to as “black box”. Once data is placed in an algorithm, it is not always clear how exactly the algorithm arrives at its prediction. It can be especially frustrating when things go wrong. A new Mechanical Engineering (MechE) course at MIT teaches students how to solve the “black box” problem, through a combination of data science and physics-based engineering.

In Class 2.C161 (Physical Systems Modeling and Design Using Machine Learning), Professor George Barbastathis demonstrates how mechanical engineers can use their unique knowledge of physical systems to control algorithms and develop predictions more precise.

“I wanted to follow 2.C161 because machine learning models are usually a ‘black box’, but this class taught us how to build a physics-informed system model so we could take a look at inside,” says Crystal Owens, a mechanic. engineering graduate student who completed the course in spring 2021.

As Chair of the Committee on the Strategic Integration of Data Science in Mechanical Engineering, Barbastathis has had many conversations with mechanical engineering students, researchers and professors to better understand the challenges and successes they have encountered using machine learning in their work.

George Barbastathis

Professor George Barbastathis teaches mechanical engineering students to use their knowledge of physical systems to develop more accurate machine learning models and algorithms. Credit: Tony Pulsone

“A comment we heard frequently was that these colleagues can see the value of data science methods for the problems they face in their research focused on mechanical engineering; yet they lack the tools to make the most of it,” says Barbastathis. “Mechanical, civil, electrical and other engineers want a fundamental understanding of data principles without having to convert to full-time data scientists or AI researchers.”

Additionally, as mechanical engineering students transition from MIT into their careers, many will one day have to manage data scientists on their teams. Barbastathis hopes to prepare these students for success with Class 2.C161.

Bridging MechE with MIT Schwartzman College of Computing

Class 2.C161 is part of the “Computing Core” of MIT Schwartzman College of Computing. The aim of these courses is to connect data science and physics-based engineering disciplines, such as mechanical engineering. Students take the course alongside course 6.C402 (Modeling with machine learning: from algorithms to applications), taught by professors of electrical engineering and computer science Regina Barzilay and Tommi Jaakkola.

Both courses are taught simultaneously during the semester, exposing students to both the fundamentals of machine learning and domain-specific applications in mechanical engineering.

In 2.C161, Barbastatthis highlights how complementary physics-based engineering and data science are. Physical laws present a number of ambiguities and unknowns, ranging from temperature and humidity to electromagnetic forces. Data science can be used to predict these physical phenomena. Meanwhile, having an understanding of physical systems helps ensure that the resulting output of an algorithm is accurate and explainable.

“What is needed is a deeper combined understanding of the associated physical phenomena and the principles of data science, especially machine learning, to bridge the gap,” adds Barbastathis. “By combining data with physical principles, the new physics-based engineering revolution is relatively immune to the ‘black box’ problem faced by other types of machine learning.”

Equipped with a working knowledge of the machine learning topics covered in class 6.C402 and a deeper understanding of how to combine data science with physics, students are tasked with developing a final project that solves a real physical system.

Develop solutions for real-world physical systems

For their final project, 2.C161 students are asked to identify a real-world problem that requires data science to resolve the inherent ambiguity of physical systems. After obtaining all relevant data, students are asked to select a machine learning method, implement the chosen solution, present and critique the results.

Topics in the last semester ranged from weather forecasts to gas flow in combustion engines, with two teams of students drawing inspiration from the ongoing Covid-19 pandemic.

Owens and his teammates, fellow graduate students Arun Krishnadas and Joshua David John Rathinaraj, set out to develop a model for the deployment of the Covid-19 vaccine.

“We developed a method to combine a neural network with a susceptible-infected-recovered (SIR) epidemiological model to create a physics-based prediction system for the spread of Covid-19 after vaccinations have started,” says Owens. .

The team took into account various unknowns, including population mobility, weather conditions and the political climate. This combined approach resulted in a prediction of the spread of Covid-19 during vaccine deployment that was more reliable than using the SIR model or a neural network alone.

Another team, including graduate student Yiwen Hu, developed a model to predict mutation rates in Covid-19, a topic that became all too relevant when the delta variant began its global spread.

“We used machine learning to predict the time-series-based mutation rate of Covid-19 and then incorporated it as an independent parameter into the prediction of pandemic dynamics to see if it could help us. to better predict the trend of Covid-19. pandemic,” Hu says.

Hu, who previously conducted research on how vibrations on coronavirus protein spikes affect infection rates, hopes to apply the physics-based machine learning approaches he learned in 2.C161 to his research on de novo protein design.

Regardless of the physical systems students tackle in their final projects, Barbastathis was careful to emphasize a unifying goal: the need to assess ethical implications in data science. While more traditional computational methods like facial or voice recognition have proven to be plagued with ethical issues, there is an opportunity to combine physical systems with machine learning in a fair and ethical way.

“We must ensure that the collection and use of data is done in a fair and inclusive way, respecting the diversity of our society and avoiding the well-known problems that computer scientists of the past have faced,” says Barbastathis.

Barbastatthis hopes that by encouraging mechanical engineering students to master both ethics and data science, they can develop reliable and ethically sound solutions and predictions for physics-based engineering challenges.

Comments are closed.