Building fusion reactors with Artificial intelligence

Introduction

Imagine combine artificial intelligence, the most powerful technology ever build by humans, and nuclear fusion reactor, with the purpose of build sustainable and clean energy source for humanity.

This is what DeepMind have done, presenting in the following paper. An incredible application of AI capabilities to help to control plasma in tomahawks fusion reactors.

In this article, I will explain you this amazing work in detail.

Fusion reactors in a nutshell

Nuclear reactions is one of the most powerful energy sources at the time, we can find fission reaction, which is basically split the nucleus of a heavy atom in two part and generating energy on the split.

The second, fusion reaction, take two atom and smash it one to another other until they become a single atom with a heavy nucleus.

When the fusion is achieved it release a tremendous amount of energy, but to achieve this, you have to push the atom with enough energy until they overcome the electromagnetic force that repelled.

This only can be achieved in extremely temperatures and pressure condition, all confine into a plasma.

The key here is that plasma, with the right configuration, shape, and conditions, you can achieve and confine a fusion reaction inside it.

Tomahawk fusion reactors designs challenge

The problem is that it is not that easy to try different plasma shapes and configurations.

Tomahawk fusion reactors are nuclear devices with torus shape, inside this reactor it produces a strong magnetic field, the work of the magnetic field is to control the plasma and change his configuration (shape)

This is made by magnetic actuator coils, basically coil controlled by a feedback system to produce different plasma configuration.

But here is the challenge, the goal is to optimize the plasma configuration, to do this you need to try a lot of different configuration but also “train” your control system to optimize it.

This is a high-dimensional issue, means that exist a ton of parameter to optimize for. This when AI comes to the rescue.

AI has a new sustainable solution

The AI-research department of DeepMind present a new architecture solution of the magnetic controller that can autonomously learn how to operate the coils and can learn the optimal plasma configuration.

The learning architecture have 3 phases, first, a designer will set some properties and parameters desired for the plasma.

In the second phase does design parameters will be passed to a deep RL algorithms, this model will start to train on a tokamak simulator, until it find the optimal policy to meet the specification from the designer.

In the third and final phase, the RL will run in a real tokamak hardware.

The designer can set from basic properties and objective to more complex configuration like plasma current, shape, and elongation.

Reinforcement learning techniques

The research team build a simulator with enough precision to describe the entire evolution of the plasma shape and current. This with the purpose of collecting data for the training of the reinforcement learning model.

The deep RL model was build to work with small amount of data since require working with the simulator it will make the computation more expensive.

To make the model faster do to the slow rate of the data the use reinforcement learning technique called posterior policy optimization

this technique is a new state-of the art reinforcement learning algorithm, but the key here is that it well suits for robust amount of hyperparameters, and it samples efficient, which mean it can learn and perform better with small amount of data.

Performance and results

The architecture presenting by the DeepMind research team demonstrate a successful control over the plasma equilibrium and also work effectively to make scientific experiment with new and more complex plasma configurations.

This demonstration open a complete new way to work with nuclear energy and AI together. Just think about the possibility in the future of energy using AI.

Conclusion

I hope you enjoy this article as much I do research, we are just touching the tip of the iceberg of artificial intelligence application in scientific and engineering world.

If you want more content like this, don’t forget to follow and comment any question that you have.

If you enjoy this content you can support it by “buy me a coffee ☕” thanks in advance.