Five years ago, Eric Aguilar was fed up.
He had worked on lidar and other sensors for years at Tesla and Google X, but the technology always seemed too expensive and, more importantly, unreliable. He replaced the lidar sensors when they broke—which was all too often, and seemingly at random—and developed complex calibration methods and maintenance routines just to keep them functioning and the cars drivable.
So, when he reached the end of his rope, he invented a more robust technology—what he calls the “most powerful micromachine ever made.”
Aguilar and his team at startup Omnitron Sensors developed new micro-electro-mechanical systems (MEMS) technology that he claims can produce more force per unit area than any other. By supplying new levels of power to micromirrors, the technology is capable of precisely steering lidar’s laser beams, even while weathering hazardous elements and the bumps and bangs of the open road. With chips under test by auto-industry customers, Omnitron is now modifying the technology to reduce the power consumed by AI data centers.
Lidar, a scanning and detection system that uses lasers to determine how far away objects are, is often adopted by self-driving cars to find obstacles and navigate. Even as the market for lidar is expected to grow by 13.6 percent annually, lidar use in the automotive industry has remained relatively stagnant in recent years, Aguilar says, in part because the technology’s lifespan is so short.
Vibration from bumpy roads and severe environmental conditions are the biggest reliability killers for automotive lidar, says Mo Li, who studies photonic systems at the University of Washington. The optical alignment within the lidar package atop self-driving cars is delicate—tremors from a poor paving job could physically alter where the mirrors sit in the housing, potentially misaligning the beam and causing the system to fail. Or temperature fluctuations could cause parts to expand or contract with the same unfortunate outcome, he explains.
Aguilar wondered which part broke most often and found the culprit to be scanners, the parts responsible for angling small mirrors that direct the laser beam out of the system’s housing. He wanted to make scanners that could withstand the tough conditions lidar faces, and silicon flexures stood out as a solution. These structures act like springs and allow for meticulous control of the mirrors within lidar systems without wearing out, as the standard metal springs do, Aguilar claims.
Designing a better chip
Aguilar hoped the new material would be the answer to the problem that plagued him, but even silicon springs didn’t make lidar systems as robust as they needed to be to withstand the elements they faced.
To make lidar even stronger, the team at Omnitron aimed to design a more powerful MEMS chip by increasing the amount of force the device can apply to control the mirrors in the lidar array. And they claim to have achieved it—their chip can exert 10 times more force per unit area on an actuator that positions a micromirror or other sensor component than the current industry standard, they say. That extra force allows for extremely valuable control in fine adjustment.
To reach this achievement, they had to dig deep—literally.
Omnitron’s micromirrors steer lidar beams and could find use in data centers.Omnitron
In this MEMS device, the mirror and its actuator are etched into a single silicon wafer. On its non-mirror end, the actuator is covered with tiny, closely spaced plates that fit between trenches in the wafer, like the interlocking teeth of two combs. To move the mirror, voltage is applied, and electrostatic forces angle the mirror into a specific position by moving the plates up and down within the trenches as the electric field pulls across the trench sidewalls.
The force that can be used to move the mirror is limited by the ratio of depth to width of the trenches, called aspect ratio. Put simply, the deeper the trenches are, the more electrostatic force can be applied to an actuator, which leads to a higher range of motion for the sensor. But fabricating deep, narrow trenches is a difficult endeavor. Overcoming this limiting factor was a must for Aguilar.
Aguilar says Omnitron was able to improve on the around 20:1 aspect ratio he notes is typical for MEMS (other experts say 30:1 or 40:1 is closer to average these days) , reaching up to 100:1 through experimentation and prototyping in small university foundries across the United States “That’s really our core breakthrough,” Aguilar says. “It was through blood, sweat, tears, and frustration that we started this company.”
The startup has secured over $800 million in letters of intent from automotive partners, Aguilar says, and is two months into an 18 month plan to prove that it can produce its chips at full demand rate.
Even after verifying production capabilities, the technology will have to face “very tough” safety testing for thousands of consecutive hours in realistic conditions, like vibrations, thermal cycles, and rain, before it can come to market, Li says.
Saving power
In the meantime, Omnitron is applying its technology to solve a different problem faced by a different industry. By 2030, AI data centers are expected to require around 945 terawatt hours to function—more than the country of Japan consumes today. The problem is “the way data moves,” Aguilar says. When data is sent from one part of the data center to another, optical signals are converted into electrical signals, rerouted, and then turned back to optical signals to be sent on their way. This process, which takes place in systems called network switches, burns huge amounts of power.
Google’s solution, called Apollo, is to keep the data packets in the form of optical signals for the duration of their travels, which yields a 40 percent power savings, the company claims. Apollo does so by using an array of mirrors to direct the data. Aguilar is planning to make the process even more efficient using dense arrays of Omnitron’s more powerful mirrors. Doing so could quadruple the amount of data each network switch could route by increasing the number of channels in each switch from 126 to 441, Aguilar says.
Omnitron is still early in its data center implementation, so it’s not yet clear to what degree this technology can really improve on Google’s Apollo. However, following a “critical design review” in mid-September, “one of the world’s top AI hyperscalers has requested our mirrors for their next generation switch,” Aguilar says. “This is proof that Omnitron solves a problem that even the biggest AI infrastructure companies can’t address in house.”
And there may be even more applications to come. Omnitron has received feelers from the defense industry, space companies, and groups interested in methane detection, says Aguilar. “It’s pretty cool seeing the people knock on our door for this because I was just focusing on lidar,” he says.