Masters of atoms 

Dr. Tess Smidt

One of humanity’s greatest feats is its increasing ability to understand and manipulate matter. Yet as far as we’ve come, from the Stone Age to the Information Age, we are far from mastering matter in its most fundamental form —  the atom. 

 

Like toddlers, puppies, and other small things, atoms are uncooperative. Putting atoms where you want them and seeing where they actually ended up is really hard. Because of this, I hung up my lab coat in graduate school to see if I could make progress on these problems using computers.

 

Given an arrangement of atoms, scientists can predict its properties using existing computational tools. The challenge is, how do we invent new, useful arrangements that are worth experimental efforts?

 

Just as written language is not all combinations of letters, spaces, and punctuation, atoms do not form random arrangements. They form recurring geometric patterns at many length scales.

 

But, the rules of this atomistic language are complex and very hard to write down. Valid geometries have to satisfy quantum mechanics, and quantum mechanics is a harsh design critic.

 

I set out to build an algorithm that can learn the language of atomic arrangements using a machine learning technique called deep learning. If this program can observe and understand the geometric patterns we've seen in experiments, it can suggest new ways of combining these patterns.

 

Deep learning algorithms and the neural networks they’re made of are top performers in translating text and identifying whether an image contains cats. However, existing deep learning algorithms are not built for handling geometry.

 

If I take your favorite molecule (such as water), and rotate it -- you and I know it’s the same thing, but a neural networks will think it is an entirely different object.

 

To fix this problem, I built a new type of neural network that can identify the same geometric pattern in any orientation or location after seeing only one example.

 

This property makes these networks useful beyond my initial goal of producing new atomic geometries. 

 

These networks can efficiently emulate expensive physics simulations because they are guaranteed to obey the symmetries of physics.

 

For example, using these networks, I can predict quantum mechanical properties of a given arrangement of atoms six orders of magnitude faster than before. In the time of this talk, I could’ve done all the supercomputer calculations I did during my entire Ph.D.

 

Right now I'm building networks to generate new atomic structures that we have never seen before, which may help guide the notorious challenge of making and measuring new materials. It’s my hope that with these tools in hand — we are well on our way to becoming masters of atoms.