June 7, 2023

machine learning brain image

When new technology meets the real world, dynamic challenges threaten to derail progress, like a self-driving car that struggles to perceive rapid changes in the environment and adjust.

The reason for that is uncertainty, and it's a challenge for all technology. Many of these ultra-smart new technologies are based on machine learning, which involves feeding data to machines that learn over time – they show improvement in performing tasks with repetition. But, in the real world, data is often limited, making it harder for machines to make accurate predictions and reactions. 

Tan Bui-Thanh, an associate professor in the Cockrell School of Engineering's Department of Aerospace Engineering and Engineering Mechanics and the Oden Institute for Computational Engineering and Sciences, has recently received several grants to develop scientific machine learning models that take into account uncertainty and use underlying governing mathematical models to fill in data gaps, the same way our brains do. The result is the ability to use limited information and physics laws to create machine learning models that are able to do as much or even faster than real-time processing and predicting than models that rely on only the governing physics or that are purely data-driven.

"In real-life applications, data is often limited, but if we can develop machine learning models that understand the physics – dynamics, physical and mathematical laws, etc. – we don't need nearly as much data to train the machines," said Bui-Thanh.

Because machine learning is still in its early days, quantifying the uncertainty in a machine learning prediction in a mathematical and physical meaningful fashion remains a critical challenge for reliable and operational machine learning models.  

The Grants: To match the potential of new technologies like scientific machine learning, experts are in the midst of understanding how they work, how to automatically determine their architecture, and how to quantify their uncertainty. The goal is to make them operational and reliable as traditional applied mathematics approach while achieving real-time capabilities.

"In order for machine learning technologies to be adopted by industries and real-life decision-making scenarios, these challenges must be overcome." Bui-Thanh said.

The three single-PI grants total more than $1 million. They are through the National Science Foundation's Office of Advanced Cyber Infrastructure, the Office of Fusion Energy Sciences' Machine Learning (ML) and Artificial Intelligence (AI) for Fusion Energy Sciences program with Los Alamos National Lab, and the Oakridge National Lab.

Why It Matters: The unpredictability of weather, for example, is a fact of life. And a big reason is a lack of data. But we do know the fundamental physics of how weather systems work, and there is plenty of available data on past weather observations (such as hurricanes).

Besides other applications such as fusion energies and fluid dynamics, Bui-Thanh’s work will combine physics knowledge from domain scientists with historical weather data to make better predictions about things like hurricane threats.

In addition to filling in the blanks in areas where data is limited, these types of models can also make improvements in situations where data isn't necessarily the problem, but time and cost represent major hurdles.

Machine learning models take major supercomputing resources and lots of time to train. Many of the most promising applications of machine learning will be challenging to implement because potential users don't have access to those computing resources or the finances to make it happen.

If new technology can reduce processing, financial and time resources needed through the use limited data sets, machine learning can become more accessible to users that lack resources and have major needs.

"Creating these new technologies is important, but so too is finding ways that allow others including researchers, engineering, and scientists to use them," Bui-Thanh said. “Part of our effort is to make scientific machine learning seamlessly accessible to broader audiences like what chatGPT could do.”