1/n Generative AI has enabled software that dreams up new code, poetry, symphonies and translates languages universally. That’s all great for the world of bits. What about for the world of atoms? We show our AI designing a lithium ion battery cathode from scratch.

2/n In the early 1930s when quantum mechanics as a picture had finally come together (Heisenberg’s uncertainty principle, entanglement, superposition and spin were all discovered), it became clear that the behavior of atoms were dictated by a different set of almost alien rules..

3/n one of the most striking realizations came from Paul Dirac, who predicted the existence of Antimatter. It became clear to him that all materials or chemical compounds, whether that’s a clump of atoms representing Advil, the next generation of solar or battery cells, have an associated many-body Schrödinger equation to them whose solutions if we could calculate them would enable us to predict a number of relevant properties. That is crazy! It means that every material imaginable, whether it’s something we discovered already such as bronze or plastics, all the way to arbitrary compounds and advanced materials we can imagine have this predictive power to them. 🤯

4/n the catch however, turned out to be that finding a solution to the many body Schrödinger equation in its exact form for relevant system sizes of dozens of atoms or more remains a *computationally intractable problem* even today despite the power of our supercomputers. They fundamentally can’t solve these equations. In order to counter act this, Kohn Sham and a number of colleagues discovered a way to approximate the solution to many body Schrödinger equations for materials we care about by introducing some clever “tricks” to solving the problem.

5/n the first major trick was to introduce the Born-Oppenheimer approximation. See, the Schrödinger equation is just the F=MA equivalent for the world of atoms. Given a set of initial + boundary conditions and leveraging the laws of conservation of energy and momentum, we can predict the evolution of a system of atoms and where they should “settle” after a given amount of time. This is important to know as it will tell us about a number of properties that directly impact the behavior of materials in our daily lives. We can predict the electrical conductivity, heat resistance, mechanical strength and optical properties of materials and compounds before even making them to a decent degree of accuracy for an increasingly large number of industry applications.

6/n back to the Born-Oppenheimer approximation…the Schrödinger equation contains two important concepts: the wave function and the Hamiltonian. The wave function encodes for the position and momentum of atoms/electrons and subatomic particles. The Hamiltonian encodes the amount of potential and kinetic energy in the system. These equations represent very complex, high dimensional partial differential equations to solve. To simplify them, Born and Oppenheimer realized that they can ignore the kinetic contributions of the nucleus of an atom since their relative velocity to electrons, which are speeding and whirring by is effectively zero. This simplifies things significantly! We only have to worry about the electronic Hamiltonian individually. The next major trick Kohn Sham et al implemented was the use of density functionals. They realized we can approximate the solution to the Schrödinger Equation by introducing functions of functions (functionals) that represent the ground state or the solution for a given Hamiltonian. Hence, Density Functional Theory (DFT) was born. DFT since it’s an approximation is always wrong to some extent. However it enables us to predict the behavior of advanced materials to a useful degree often enough that it’s one of the first things you do as a materials scientist after literature review and initial experimental testing. You want to know if the material you’re going to spend millions of dollars and months or years on will behave the way you expect it to. Maybe you have a new alloy that will disrupt the automotive industry in mind, or a new pharmaceutical with specific structure-activity properties that enable the cure for neurodegenerative diseases. It’s all the same type of problem to the computer.

7/n At @QuantumGenMat we train our AI on large datasets of DFT, molecular dynamics (the motion and interactions of atoms and their electrons) as well as physical real world data to make useful predictions about new materials. In this case, we wanted to do a simple test of seeing whether we could predict a delithiated structure associated to the lithium ion nickel manganese cobalt (NMC) family of cathodes. Conventionally, you would need to have started with some known structure and iterated from there and it would have taken thousands of times longer (plus costed more) to predict. Whereas we started with an arbitrary combination of atoms and predicted a stable structure in seconds from there.* *Stable is a relative metric and does not mean it is stable in the real world. This is where the bulk of our work goes into as being able to reliably, consistently and efficiently predict novel stable structures that can be synthesized in the lab with AI is still very much an unsolved and outstanding task. It’s still mind blowing to get a glimpse of what AI can do today for materials discovery. The future is exotic. 🫡⚛️

How are you dealing with compositional disorder? The NMC cathode materials you choose are solid solutions with disordered metals. How are you simulating these with DFT?

While this example doesn’t deal with disorder (not in the scope of this work) you could probably introduce disorder via setting up bigger supercells. Please note that this is not the best use case for these generative models anyways, just a PoC. More relevant areas than NMC

My pov is that compositional disorder is so important that it can't be ignored when calculating new Inorganic materials beyond very simple compounds. There are ways to tackle this like with large supercells but this can't be an afterthought - needs to be fully integrated otherwise we see the problems experienced by Deepmind where none of the results could be relied upon.

Why is that AI and not just an algorithm? It looks like simulation algos I've been seeing for 3 decades.

Recap: it's nothing, but it could be something and think of the future, it could be even more! 😂🤣

The birth of Cathodisicm.

Ight... hear me out. We make a new type of vehicle & battery at the same time.

generative evolution