A chatbot advises you to put glue on your pizza. Or eat one rock each day. Or says that your neighbor might be an alien. Today, more than 50% of Americans have used artificial intelligence, or AI, tools, like ChatGPT with results that have been both rewarding and risky.
Scientists, too, are making increasing use of a form of AI called scientific machine learning to spur rapid advancements in fields ranging from health care to material science. But there, it is even more important that AI delivers trustworthy results.
“In science, the stakes are higher,” says Gunther H. Weber, staff scientist in the Scientific Data Division in the Computing Sciences Area at the Lawrence Berkeley National Laboratory, or LBNL. “If you point your million-dollar telescope at the wrong point in the sky, the results are much worse than if a generative AI tool creates a bad photo.”
Scientific machine learning is a branch of AI that tackles complex challenges by blending traditional scientific principles — such as those from physics or chemistry — with machine learning, enabling computers to identify patterns in data. This method is particularly valuable in areas where data is limited, and the problems are difficult to model.
But scientists must be sure that these AI systems are serving up reliable results and not the high-tech equivalent of pizza with glue.
To address these issues, Weber and the team at LBNL have established a collaboration with Arizona State University. Led by Ross Maciejewski, director of the School of Computing and Augmented Intelligence, part of the Ira A. Fulton Schools of Engineering at ASU, the researchers are developing advanced data visualization tools to help experts understand scientific machine learning models.
They have created Landscaper, an open-source Python-based suite of software and tools that scientists can use like night-vision goggles for a pitch-dark AI jungle, allowing them to see what’s going on and avoid hidden traps.
Through a broader collaboration known as LossLens, the team is also leading student research experiences designed to train the next generation of computer and data scientists to address emerging challenges.
Read the full story on Full Circle.