Aerial view of power plant.
An arrow pointing leftHome

This startup wants to use AI to teach power plants and steel factories how to run themselves

  • Mike Pearl

Hard hats, metal lunch boxes and a computer in charge of it all?

In the first two Terminator films, eliminating the titular AI killing machines came down to two climactic showdowns in industrial settings: a machine parts factory and a steelworks, respectively. One interpretation of this recurring motif is that the only locations in America where the concentrations of raw physical power are sufficient to stop the computerized terror of our dystopian future are in its industrial core.

Phaidra, a Seattle-based AI start-up that just received $25 million in venture capital, would like to flip that script upside-down. According to its website, the company “creates self-learning control systems for mission-critical industries,” a task that would enable “efficiency, sustainability and control stability to automatically improve.”

In other words: put the AI in charge of the raw physical power, and maybe that will make our future less dystopian.

Phaidra’s leadership is truly a mix of big industry and Silicon Valley brains. The co-founder and CEO is Jim Gao, who worked at Alphabet’s DeepMind AI research hub, and co-founder and CTO Vedavyas Panneershelvam is also a DeepMind alum. However, a third co-founder, Katherine Hoffman, worked in the HVAC and defense industries before moving to AI, and company president Robert Locke spent 13 years at an industrial supply company called Johnson Controls.

Marketing their business solutions in the old-fashioned world of heavy industry — clients reportedly include power plants and paper mills — sounds challenging, since it seems these companies often do have some sort of computerized power optimization in place. Hoffman ReadAItold Bloomberg last week that U.S. industrial companies are optimizing their energy use with “what’s been around since the 1950s,” and that “these industrial systems are incredibly difficult to run on a good day.”

The AI involved is drawn from deep reinforcement learning — the kind that can be used to “solve” a game like Go or checkers. In fact, a 2015 paper co-written by Phaidra co-founder Vedavyas Panneershelvam explores the use of deep reinforcement learning — a version of a deep Q network, specifically — to train an AI to play Atari 2600 video games.

Phaidra’s clients are promised a software-only solution, so apparently Phaidra simply pairs with companies’ existing software used for these purposes. Phaidra software, then, can be thought of as the hands operating the video game controller, except instead of getting a high score for killing bad guys, it gets a high score for optimizing energy use.

Phaidra’s precision movement of the joystick, if you will, ideally hits just the right temperature for a facility in any given context — be it a vaccine factory, or a steelworks. Phaidra also has to watch the system’s “life bar” carefully — which is to say: it knows when equipment is going to start to lose efficacy due to a problem with temperature. Intelligently heating and cooling an industrial plant in this way supposedly uses up to 30% less power.

Phaidra demurred when Bloomberg asked what it charges clients for its software, and also when GeekWire asked if it was a profitable company, so there’s still a ways to go before Phaidra is a mainstream SaaS brand, but for what it’s worth, it’s a forward-thinking one, assuming the wind-down of greenhouse gas emissions from energy use becomes part of everyday reality for business people.

Phaidra is betting that it will. “Energy efficiency is no longer a nice-to-have, Gao told. “It is now business critical.”