DeepMind scientists say they trained an A.I. to control a nuclear fusion reactor
Scientists at DeepMind have found another real-world application for their artificial intelligence software.
The London-based AI lab, which is owned by Alphabet, announced Wednesday that it has trained an AI system to control and sculpt a superheated plasma inside a nuclear fusion reactor.
Nuclear fusion, a process that powers the stars of the universe, involves smashing and fusing hydrogen, which is a common element of seawater.
The process, which releases vast amounts of energy, has been touted as a potentially limitless source of clean energy, but a number of technical challenges still need to be overcome.
Here on Earth, scientists use tokamaks — doughnut-shaped vessels surrounded by electromagnetic coils — to try to recreate the nuclear fusion reactions that occur in outer space.
The magnets in these tokamaks are used to “contain” the volatile hydrogen plasma, which is hotter than the core of the sun. Controlling the magnetic coils currently requires multiple layers of complex control systems.
Martin Riedmiller, control team lead at DeepMind, and co-author of a paper published in the journal Nature on Wednesday, told CNBC that it's “a really complex” control problem.
Together with the Swiss Plasma Center at EPFL, a university in Lausanne, Switzerland, DeepMind said it has developed a reinforcement learning AI system that can control the magnets and change their voltage thousands of times per second.
Reinforcement learning, an AI training technique that DeepMind is particularly bullish on, involves programming an AI to take certain actions in order to maximize its chance of earning a reward in a particular situation. In other words, the algorithm “learns” to complete a task by seeking out these preprogrammed rewards.
DeepMind's unnamed AI, developed on a virtual simulator, has been used around 100 times on a tokamak at the Swiss Plasma Center known as the Variable Configuration Tokamak. It controlled the magnets in the tokamak for two seconds, which is the maximum amount of time the reactor can run before it overheats.
Roughly 10-20 people from DeepMind worked on the AI system together with around 5-10 people from EPFL.
“Fusion as I see it is one of the most fundamental energy sources that we have in the entire universe,” Federico Felici, a research Scientist at the Swiss Plasma Center, told CNBC. “Once we actually master this technology...it's a huge achievement because you will be able to have near be limitless energy for the foreseeable.”
Damien Ernst, a professor at Universite de Liege who was not involved in the work, hailed the research as one of the most important applications of reinforcement learning so far, adding that it “could dramatically accelerate the development of fusion reactors and, ultimately, our ability to fight global change.”
DeepMind set out to crack artificial general intelligence, which is often referred to as the holy grail of AI.
The company started by developing AI systems that could master games like chess and Go. Now, it wants to apply more of its technology to real world applications and science.
While Google has found uses for DeepMind's AI, its technology has not been widely applied elsewhere.
DeepMind CEO Demis Hassabis said in a statement that the company has demonstrated AI's potential to accelerate scientific progress and open new avenues of research across biology, chemistry, mathematics and now physics.
DeepMind employs about 1,000 people worldwide, including some of the world's leading AI research scientists, who can command annual salaries of more than $1 million. These top people, who often have Ph.D.s from the likes of Oxford, Cambridge, Stanford and MIT, can command this sort of money because they're also sought after by Big Tech companies like Facebook, Apple, Amazon and Microsoft.
DeepMind 的科学家们为他们的人工智能软件找到了另一个现实世界的应用程序。 Alphabet 旗下位于伦敦的人工智能实验室周三宣布，它已经训练了一个人工智能系统来控制和雕刻核聚变反应堆内的过热等离子体。
核聚变是一个为宇宙恒星提供动力的过程，涉及粉碎和融合氢，这是海水的常见元素。
这个释放大量能量的过程被吹捧为一种潜在的无限清洁能源，但仍有许多技术挑战需要克服。
在地球上，科学家们使用托卡马克——被电磁线圈包围的甜甜圈状容器——试图重现外太空发生的核聚变反应。
这些托卡马克中的磁铁用于“容纳”比太阳核心更热的挥发性氢等离子体。目前控制磁线圈需要多层复杂的控制系统。
DeepMind 控制团队负责人、周三在 Nature 杂志上发表的一篇论文的合著者 Martin Riedmiller 告诉 CNBC，这是一个“非常复杂”的控制问题。
DeepMind 与位于瑞士洛桑的一所大学 EPFL 的瑞士等离子体中心合作开发了一种强化学习 AI 系统，该系统可以控制磁铁并每秒改变数千次电压。
强化学习是 DeepMind 特别看好的一种人工智能训练技术，它涉及对人工智能进行编程以采取某些行动，以最大限度地提高其在特定情况下获得奖励的机会。换句话说，算法通过寻找这些预编程的奖励来“学习”完成任务。
DeepMind 的未命名人工智能是在虚拟模拟器上开发的，已在瑞士等离子中心的一个名为可变配置托卡马克的托卡马克装置上使用了大约 100 次。它控制托卡马克中的磁铁两秒钟，这是反应堆在过热之前可以运行的最长时间。
来自 DeepMind 的大约 10-20 人与来自 EPFL 的大约 5-10 人一起研究 AI 系统。
“在我看来，聚变是我们在整个宇宙中拥有的最基本的能源之一，”瑞士等离子体中心的研究科学家 Federico Felici 告诉 CNBC。 “一旦我们真正掌握了这项技术……这是一项巨大的成就，因为在可预见的情况下，你将能够拥有近乎无限的能量。”
列日大学教授 Damien Ernst 没有参与这项工作，他称赞这项研究是迄今为止强化学习最重要的应用之一，并补充说它“可以极大地加速聚变反应堆的发展，并最终加速我们的应对全球变化的能力。”
DeepMind 着手破解通用人工智能，这通常被称为人工智能的圣杯。
该公司从开发可以掌握国际象棋和围棋等游戏的人工智能系统开始。现在，它希望将更多技术应用于现实世界的应用和科学。
虽然谷歌已经找到了 DeepMind 人工智能的用途，但其技术并未在其他地方广泛应用。
DeepMind 首席执行官 Demis Hassabis 在一份声明中表示，该公司已经展示了人工智能在加速科学进步和开辟生物学、化学、数学和现在物理学的新研究途径方面的潜力。
DeepMind 在全球拥有约 1000 名员工，其中包括一些世界领先的 AI 研究科学家，他们的年薪可超过 100 万美元。这些顶尖人物通常拥有牛津、剑桥、斯坦福和麻省理工学院的博士学位，他们可以赚到这种钱，因为他们也受到 Facebook、苹果、亚马逊和微软等大型科技公司的追捧。
