CRYSTALS CAN do all sorts of things, some more useful than others. They can separate the gullible from their money in New Age healing shops. But they can also serve as the light-harvesting layer in a solar panel, catalyse industrial reactions to make things like ammonia and nitric acid, and form the silicon used in microchips. That diversity arises from the fact that "crystal" refers to a huge family of compounds, united only by having an atomic structure made of repeating units-the 3D equivalent of tessellating tiles.
水晶能做各种各样的事情，有些比其他的更有用。它们可以在新时代疗愈商店中骗走傻子的钱。但它们也可以作为太阳能电池中的光收集层，催化工业反应制造氨和硝酸等物质，并形成微芯片中使用的硅。这种多样性来自于“水晶”这个词指的是一个庞大的化合物家族，它们唯一的共同点是拥有由重复单元组成的原子结构，即三维版本的镶嵌瓷砖。
Just how huge is highlighted by a paper published in Nature by Google DeepMind, an artificial-intelligence company. Scientists know of about 48,000 different crystals, each with a different chemical recipe. DeepMind has created a machine-learning tool called GNOME (Graph Networks for Materials Exploration) that can use existing libraries of chemical structures to predict new ones. It came up with 2.2m crystal structures, each new to science. To check the machine's predictions, DeepMind collaborated on a second paper, also published in Nature, with researchers at the University of California, Berkeley. They chose 58 of the predicted compounds and were able to synthesise 41 of them in a little over two weeks. The team at Deep Mind say more than 700 other crystals have been produced by other groups since they began preparing their paper.
这一巨大规模在Google DeepMind，一家人工智能公司，发表在《自然》杂志上的一篇论文中得到了凸显。科学家们知道大约有48,000种不同的水晶，每一种都有不同的化学成分。DeepMind创建了一种名为GNOME（用于材料探索的图网络）的机器学习工具，可以使用现有的化学结构库来预测新的化合物。它提出了2.2百万个水晶结构，每一种在科学上都是新的。为了验证机器的预测，DeepMind与加利福尼亚大学伯克利分校的研究人员合作，发表了第二篇论文，他们选择了其中的58种预测化合物，并在短短两周多的时间内合成了其中的41种。DeepMind团队表示，自他们开始准备论文以来，其他研究小组已经产生了700多种水晶。
To help any other laboratories keen to investigate the computer's bounty, the firm has made public a subset of what they think should be the 381,000 most stable structures. Among them are many thou sands of crystals with structures potentially amenable to superconductivity, in which electrical currents flow with zero resistance, and several hundred potential conductors of lithium ions that could find a use in batteries. In both cases Deep Mind's work has increased the total number of candidate materials known to researchers tens of times over.
为了帮助其他实验室研究计算机的成果，该公司已经公开了他们认为应该是381,000个最稳定结构的子集。其中包括许多结构有望实现超导性的水晶，电流在其中零电阻流动，并且数百种可能用于电池的锂离子导体。在这两种情况下，DeepMind的工作将研究人员已知的候选材料总数增加了数十倍。
Aron Walsh, a materials scientist at Imperial College London who was not involved in the research, says DeepMind's work is impressive. But "this is the start of the exploration rather than the end," he says, noting that the machine has only scratched the surface of what might be possible. In a recent paper of his own he tried to calculate how many stable crystals incorporating four chemical elements (so called quaternaries) might be potentially manufacturable. He wound up with a conservative estimate of 32trn. For its part, GNOME looked only at crystals that form under relatively low temperatures and pressures. And crystals are only one subset of a universe of materials that includes everything from amorphous solids such as glass through to gases, gels and liquids. Whether any of DeepMind's 2.2m new crystals will be useful remains to be seen. Even if they do not, the techniques used to make the predictions could be valuable. Besides suggesting new crystals, AI may also shed light on as-yet-unknown rules that govern how they form.
伦敦帝国理工学院的材料科学家Aron Walsh在研究中没有参与，但他表示DeepMind的工作令人印象深刻。但他说：“这只是探索的开始，而不是结束”，指出机器只是刚刚触及了可能性的表面。在他自己的最近一篇论文中，他试图计算有四种化学元素（所谓的四元体）组成的稳定水晶可能有多少是可制造的。他得出了一个保守的估计，为32万亿。GNOME只考虑在相对低温低压下形成的水晶。并且水晶只是一个包括从非晶固体如玻璃到气体、凝胶和液体等一切材料的宇宙中的一个子集。DeepMind的2.2百万个新水晶中是否有用仍有待观察。即使它们没有用，用于进行预测的技术也可能很有价值。除了提出新的水晶外，人工智能还可能揭示迄今为止尚未知晓的规则，这些规则指导水晶的形成。
Ekin Dogus Cubuk at DeepMind highlights one such finding. Previously, he says, crystals made from six elements, called senaries, were thought to be vanishingly rare. But DeepMind's AI found around 3,200 in its sample of 381,000 stable compounds. A better understanding of how crystals form, and what sorts are possible, might also save scientists curious to test how the 2.2m new materials behave from the tedious task of synthesising each one of them by hand.
DeepMind的Ekin Dogus Cubuk强调了一个这样的发现。他说，以前认为由六种元素制成的水晶，称为六元体，被认为是极其罕见的。但DeepMind的人工智能在其381,000个稳定化合物样本中找到了大约3,200个。对水晶形成方式的更好理解以及可能存在的各种类型可能也会使科学家在测试这2.2百万种新材料的行为时免受手工合成每一种的繁琐任务。
