Artificial Intelligence Goes Nuclear
Impact on the Energy Balance isn’t Obvious
By Matthew L. Wald
Conventional wisdom is that artificial intelligence will demand vast new supplies of electricity.
But technologies hardly ever operate in only one direction, and this assessment may be backwards. It turns out that A.I. could also make energy cheaper and more plentiful. It will certainly cut both ways.
This shouldn’t be surprising. Technologies that at first glance seem to consume a lot of energy often save it too, with a net effect that is hard to determine. Those data centers that consume perhaps 2 percent of United States electricity production also power the websites that let me get a dozen widgets delivered to my door, saving me and my car a trip to the hardware store.
And the data centers may not require nearly as much energy as a straight-line projection would suggest.
In A.I.’s case, it’s likely to make energy production easier, especially nuclear energy production.
There are challenges though. One is that the Nuclear Regulatory Commission says it is committed to recognizing A.I. and allowing it to make operations more efficient at licensees, and at the commission itself, but it is still working on the rules under which this can happen. But first, the commission must have an A.I.-literate staff, which it is now seeking to develop.
Another challenge is public confidence. The technology is famous for telling a New York Times reporter in a chat that he should divorce his wife because she doesn’t really love him. There is a vague but pervasive anxiety about “the singularity,” when A.I. takes over the world.
But this is overblown, at least from what we can see now. A.I. is a tool that will help us do the kinds of things we do already, but with less effort. And the industry already has a toe in the water.
Optimizing Reactor Operations
Consider, for example, the core of a Boiling Water Reactor. There is a cartoon-like schematic of how it works, and there is a much more complex reality. The reality turns out to be very hard to optimize, but A.I. is helping.
Here’s the simple version:
On the left, the fuel, in the form of ceramic pellets of enriched uranium, arranged into rods that are bundled together, sustains a chain reaction that generates heat. This boils the water in the core, which is used to spin a turbine (on the right), which turns a generator and makes electricity. The steam is condensed back into water and pumped back to the core for re-heating.
The water in this system is not just for heat transfer; it is also a moderator, controlling the speed of the neutrons, the sub-atomic particles that sustain the reaction. It slows them down to a speed that is most likely to cause another fission when the neutron hits another uranium nucleus.
And every 18 months or two years, operators shut the reactor down and remove the oldest 1/3 of fuel assemblies in the core. The remaining assemblies are shuffled to new locations as needed and new assemblies are installed in the open locations. Then the reactor runs another 18 months or two years. After three cycles, around 5-6 years, the fuel is removed.
But that’s the simplified version. The reality is more complex. For starters, the ability of the water to slow down the neutrons depends on the water’s density. But density isn’t constant over the height of the fuel; the water at the top has more steam bubbles in it and isn’t as effective a moderator. And the operators may insert control rods at the bottom of the core to moderate the power output. So, sometimes the fuel at the bottom of the core can undergo more fission than fuel at the top, and sometimes it’s the other way around.
And the engineer who designs the core must keep in mind that the objective is a lot of heat without exceeding temperature limits at any spot. Heating by fission can be uneven, and this is more complicated than defrosting a frozen dinner in the microwave or putting something thick on the barbeque.
The core designer has tools, though. The designer can specify different levels of enrichment in various parts of the core in a way that keeps temperatures close, but not over, the limit. Another objective is to be able to keep the reactor at full power for long periods and come close to using up all the fuel at the end of the third cycle.
But it doesn’t always work that way. Sometimes after two cycles, the reactor operators realize that there are assemblies that are scheduled to run another cycle but don’t have enough un-fissioned uranium in them to last another year or 18 months. So, the plant owner may remove that fuel, which isn’t quite exhausted, and thus has a substantial amount of unused uranium still in it. (This is akin to replacing a flashlight battery on a fixed schedule, which turns out to be at a time when the battery still had hours of use left in it.)
But A.I., it turns out, may be better than humans at designing a core in which the fuel lasts as long as it’s supposed to but ends the third cycle with nothing left. A.I. can try out different pattens of enrichment within the core and calculate what the power distribution will be, what the temperatures will be, and how the fuel will be consumed. If the solution doesn’t look good, it can learn from result, vary the recipe and try again A company in West Lafayette, Indiana, Blue Wave AI Labs, says the sixteen reactors that use its software have saved more than 100 fuel assemblies.
Along with saving money, this reduces the volume of spent fuel. Spent fuel management is secure and safe, but for the time being, used fuel is sealed into steel casks, in a dry, inert environment. This is a cumbersome process, and reducing the number of casks needed would be a step forward.
Giving Smarts to Robots
People think of smart robots as something for emergencies, but they can also relieve humans of mundane tasks like taking inventory of spare parts or monitoring plant equipment.
One problem is that unmanned vehicles generally use GPS for navigation, but that doesn’t work indoors, and especially not in the tight spaces in a nuclear power plant, which have thick concrete walls, often with steel reinforcing rods embedded.
With that in mind, the Energy Department’s program to extend the life and improve the efficiency of current-generation reactors developed a system for posting QR codes that act as navigational beacons. A drone’s camera cannot only recognize the code, but with the aid of A.I., can analyze the camera’s position relative to the QR code, based on the shape of the image: Is it straight ahead, or to the right, the left, above or below? If more than one QR code is visible, the drone can triangulate.
The drone, using cues supplied by the QR code, can accomplish tasks like taking pictures of a gauge. They can also patrol like security guards or look for smoke or fire. They can file “condition reports.” And on the receiving end, A.I. can help categorize reports from A.I.-driven robots in the field, and from humans, to assess the severity of the problem, decide whom to assign to fix it, check the spare parts inventory, and order more components as needed. It can also track common issues.
The Energy Department says such a computer can predict the decisions that a human would make.
Other applications are in their nascent stages, but there is some reason to believe that A.I will be well suited to the coming generation of small reactors. One company, Atomic Canyon, is using an Energy Department supercomputer at Oak Ridge, the world’s fastest, to develop an open-source nuclear A.I. model for other companies to use.
X-energy, for example, is developing a high-temperature gas-graphite reactor, the first of which will be built at a Dow chemical plant in Texas, where it will displace natural gas in making steam for industrial uses. But these reactors will be small, relative to the behemoths operating today. To run small reactors cost-effectively, X-energy plans regional support centers that will stock spare parts, maintain a cadre of technicians with specialized skills, train operators and conduct other centralized functions. Today, some utilities that operate fleets of reactors have centralized some of those functions, but others are done at each individual plant site. X-energy’s plan is that A.I. will analyze operational data from reactors in different locations, and dispatch spare parts, specialized equipment and specialized workers from the support centers, as needed, to optimize maintenance and operations.
And some nuclear industry uses of A.I. can be borrowed from other fields, or at least informed by their experience. A challenge for extending the life of current reactors is that neutrons escaping the reactor core hit concrete structures. One of the materials frequently used in concrete aggregates (the pebbles that are mixed into turn cement into concrete) is quartz, and when quartz is struck by a neutron, it breaks down.
There is a process for non-destructive evaluation of those concrete structures, akin to a medical CAT scan, which produces an image in which the damaged quartz shows up differently. But the difference is faint, and a human evaluating the scans may miss it. A.I. can do the job according to researchers.
But while A.I. may help nuclear, its net effect on energy and on carbon isn’t clear because it will also be employed in other energy fields. For example, it can analyze 3D seismic data, which the oil and gas industry use to find underground deposits and increase the chance that a drill bit will find a reservoir. It can also predict what maintenance tasks are needed, for various levels of the industry, including drilling and refining. And anything that makes it easier to discover, produce and refine fossil fuels will make them more of a market success. Last year the Boston Consulting Group predicted that use of A.I. could cut emissions by 5 to 10 percent.
It is fair to say that we are venturing into uncharted territory. A recent report from the Department of Energy says the technology has many advantages for energy, but also that it creates new vulnerabilities. Saboteurs, for example, could feed it bad data “so that a model develops an incorrect conception of what ‘normal operations’ look like.” Or A.I. could be used to plan attack and to seek out high-value targets.
But new technology is almost always uncharted territory. And the application of A.I. to nuclear operations holds promise for increasing energy abundance. So the use of A.I. to optimize the design of new reactors, and anything that helps bring forth a new generation of advanced reactors, will be a plus in the energy equation. We live in an age when computers are increasingly applied to increase efficiency, under the hood of our cars, in the thermostat on the living room wall, in electric grid control centers, and in a thousand other settings. Nuclear reactors can join the list.
Let’s not add to the AI fluff. AI will certainly be useful for a lot of things, but a lot of the stuff mentioned in this article could be done with standard ERP systems and technology from the early 2000s. All that would be required is to bring the plants up to modern control standards instead of using INC systems from 1970…. That won’t stop the consultants bringing nonsensical white papers with techy sounding lingo- only engineers can do that.
Fair to point out that AI may improve efficiencies and there is no crystal ball. Skeptical that AI overall will reduce power demand - at best will help minimize it's impact.. The trajectory for humankind indicates more energy will be needed. With much of the world aspiring to a better life, increased energy consumption is a given barring some catastrophe. Let's at least hope AI will help with efficiency and reducing development times.