How Is AI Used in Engineering and Materials Science?

AI is already helping engineers in real, specific ways: machine learning has pointed chemists to a new capacitor ceramic, self-driving labs have grown better semiconductor films and crystals, a robot has served juice in a mall for a full shift, and phone and dashcam software has helped blind walkers and flood-prone neighborhoods. None of it is in everyday products yet. These are the cases we have reported and checked, each a careful step rather than a finished tool.

The people who stand to gain range from engineers designing phones and power chips to a blind walker heading for the bus stop and a family in a basement apartment on a block that floods without anyone noticing.

Finding and making new materials faster

Phones, radios and satellites are packed with tiny capacitors, and room inside them is always scarce. An international team, most of them in China, used machine learning to score a virtual library of 132,651 entries and find a temperature-stable ceramic, then built prototype chip capacitors from it. One took up about 85 percent less volume than a comparable commercial part, a comparison the authors made themselves. These are single-layer lab prototypes with no production timeline, and two co-authors are affiliated with a state key laboratory at Guangdong Fenghua Advanced Technology Holding Co. Read how machine learning narrowed the search for a smaller capacitor ceramic.

Researchers at NTT in Japan built one that grew gallium oxide films, promising for power electronics, and reached a disorder score of 182 meV in 56 runs, which the authors call the lowest reported for sputtered films. Then they pulled a simple rule from its data. People following it by hand reached 163 meV, though they needed 85 runs to get there. No power devices were built yet. See how a self-driving lab handed engineers a recipe they can follow.

Growing light-emitting crystals has long leaned on an expert’s instinct. A team at the Chinese Academy of Sciences’ Institute of Semiconductors and Zhejiang University built SemiEpi, which watches the crystal surface as it grows and corrects the conditions on the fly. Aiming at 1240 nm, a wavelength suited to gas sensing, it produced quantum dots at 1235 nm that were brighter and purer in color than a conventional sample. A person still handled some steps by hand, and no finished sensors exist yet. Read how SemiEpi takes some guesswork off crystal growers.

Robots steady enough to trust with real work

A robot that fails one time in twenty is no help behind a counter. Researchers from Shanghai Jiao Tong University, Tsinghua University, the Shanghai Qi Zhi Institute and other institutions trained robots with a method called RL-100, which starts from human demonstrations and improves by practice. In the lab the robots succeeded in 1,000 of 1,000 evaluated trials on eight tasks, under the authors’ own protocol, and a juicing robot served mall shoppers for about seven hours without a failure. That is one shift on one task, with no commercial deployment named. Follow the juice counter robot’s first long shift.

Helping people get around and stay safe

A Harvard team led by Raymond Liu and Patrick Slade built Mobilio, a phone app that uses a computer vision model to find sidewalks and crosswalks and steers people with beeps from the left or right. In tests with 14 volunteers, they finished an outdoor route 13 percent faster than with Google Maps and a cane, with fewer contacts with obstacles. It is a small study, Harvard is pursuing commercialization, and the app is not downloadable yet. Read about the phone app built to help blind walkers travel on their own.

Flooding that nobody reports can go unfixed. Researchers at Cornell Tech and UC Berkeley built BayFlood, which asked an open-source vision-language model whether each of 926,212 dashcam photos from one New York storm showed more than a foot of standing water, then estimated flooding tract by tract. It flagged areas home to about 1.05 million people as high risk, including about 100,150 that no sensor, complaint or stormwater map had marked. Its flood calls were right about two thirds of the time, and it ran after the storm, not live. See how dashcam photos could show cities their hidden flooding.

How close is this to everyday life?

All of these sit in the middle distance: past the idea stage, short of something you can use today. Roughly:

• Closest to real decisions: the flood maps. The team suggested 25 new sensor locations to New York’s FloodNet, which expressed interest in the data as one input among others.
• Working with real people, not yet in their hands: the Mobilio app, tested with 14 volunteers around Boston, and the juice robot, tested for one shift in a real mall.
• Proven in the lab, devices still to come: the capacitor ceramic, the gallium oxide films and the SemiEpi quantum dots. Each still needs to be built into working parts and shown at scale.

In several of these stories the software did not replace people. It narrowed a huge search, read a crystal as it grew, or turned its own data into a rule a person can carry to the next machine.

The next time a blind walker reaches a bus stop on her own, or a city puts a flood sensor on a forgotten block, it may trace back to work like this. As these projects move from the lab toward people, you can follow each new step on our engineering page.