You may have seen some stories recently about another breakthrough for AI/ML approaches in chemistry, this one around a technique called DFT. So what the heck is DFT, and what did this new work accomplish?
Up front, I have to state that I Am Not a Computational Chemist, and I have never done any of this stuff with my own hands. But when that ever slow anyone down? I have been watching it ove the years and have occasionally been a customer for it, so I can comment from that perspective. The first thing to do is talk about what Density Functional Theory is, and why people care about it. In short, it's a way to calculate the electronic properties of molecules (the ones that depend on the distribution of their electrons in the space around them). DFT is a shortcut, an approximation, but there's no other way to calculate these things on useful-sized molecules other than by making some assumptions and taking some shortcuts. Exact solutions for many-body problems like this in quantum mechanics are not feasible - you end up with N electrons, each with 3N spatial coordinates and all interacting with each other at the same time.
In DFT, as the name implies, you're using "functionals", that is, functions of functions. The underlying functions are naturally those for spatial electron density, but how you make functions out of those is where the approximations (and the computational advantages) come in. There are several things to consider. The way that one electron's behavior is influenced by the others in a system is called the "correlation energy", and finding better approximations to that has been a big part of getting DFT to work more accurately. Another improvement has been in handling the "exchange term", which is a quantum effect between identical particles. The wave functions of particles like electrons can either stay the same or flip signs when they exchange (both the spatial and spin parts have to flip), and this causes the expected distance between two electrons to be greater than you'd predict otherwise. The exact functionals for those two (correlation and exchange) aren't known, and it's quite possible that there may not be any, but there have been increasingly good methods to sneak up on useful answers.
This may all sound pretty obscure, but Walter Kohn (a key mover in this field) won a Nobel Prize for it in 1998. It's important because you can predict quite a bit about a molecule's behavior or the behavior of materials by using DFT, and these include materials like semiconductors and other electronics, the behavior of crystals in general (such as their mechanical, magnetic, and conductance properties), prediction of things like melting points, viscosity, surface tension, interactions of small molecules with surfaces (including things like drugs binding to proteins), and many, many more. DFT (in its many forms) can give really useful answers in many cases, but it has several well-known limitations and blind spots. Doing these calculations for a material in the presence of an external magnetic field is not much fun, for one thing, and situations that involve important contributions from dispersion forces are a weakness, too, one that limits DFT's power in doing many intermolecular binding calculations.
This latest work gets at another particular limitation of DFT, one that can be illustrated by good ol' sodium chloride. In reality, a single neutral sodium atom separated from a single neutral chlorine atom has no electron transfer between the two (that only happens when they get closer, and then it happens big-time, with the formation of a Na+ and Cl- ion pair). But DFT tends to predict some "fractional" electron transfer even when the two atoms are separated, assigning a pretty-good probability to that transfer and then giving you the average. But that's not real, and it can be a significant source of error. The new paper from the DeepMind team uses several large data sets (along with some concocted data with fractional charges and spins) to train a system to handle these situations better. The functional that they end up with is called DM21, and it does a significantly better job on main-group elements than the exisiting DFT functionals. That's especially true for problems that are known to involve electron transfer (or the lack of it!), but DM21 gives improved results on some other calculations where that effect is less obvious, too. This is an advance in itself, and it also points the way to further advances - as the authors note, this work relied on experimental data and on computational constraints, and both of these can be further improved.
If you've seen anyone talking about how computational chemistry or materials science has now been solved by AI, you can discount that stuff - the paper itself claims no such thing. For example, while it handles main-group elements more accurately, that likely isn't true for transition metals, or indeed for many more complex solids and liquids. Some of the constraints that DM21 uses to get better answers might break down as you go up to larger multi-atom systems like whole crystals. DeepMind set out to improve the situation around a specific defect in current DFT work, and they did that very impressively, but they did not suddenly wave away all the technique's limitations. But the success here points the way to applying similar AI approaches to the other DFT problems - if you feed these algorithms a good-sized set of reliable data from these other sorts of calculations, and with particular attention to examples where DFT diverges from experimental data, you might well make similar progress on dispersion forces and so on. Remember, we're looking for workaround approximations in ever finer degrees rather than some sort of exact mathematical insight, and the search for these is well suited to a machine-learning attack.
About the author
