AI solves problem in seconds when quantum chemistry takes days

In the middle of each chemical reaction, as molecules crash together to form something new, there is a fleeting moment when energy peaks and the course of the reaction becomes inevitable.

Called the “transition state”, this moment in time is so tiny that it can barely be recorded. But it decides the products of the reaction. Knowing the transition state helps chemists to make new materials, fuels and pharmaceuticals. It’s also important for studying biological reactions, from medical sciences through to figuring out how life first evolved.

Understing the transition state is a time-consuming process which relies on quantum chemistry. But US researchers believe they’ve found a faster route with AI.

“The transition state helps to determine the likelihood of a chemical transformation happening,” explains Associate Professor Heather Kulik, from MIT,  and senior author on a paper published in Nature Computational Science.

“If we have a lot of something that we don’t want, like carbon dioxide, and we’d like to convert it to a useful fuel like methanol, the transition state and how favourable that is determines how likely we are to get from the reactant to the product.”

At the moment, the best way to find a transition state is a quantum method called density functional theory. This can sometimes take days to calculate one transition state.

Chemists have been toying with machine learning to avoid this process. For now, this is hampered by the fact that molecules crash together at all sorts of angles and orientations during a reaction. Computer models have to develop an entirely new process to calculate each orientation.

Related Article

Core Sciences

What does a poisonous alcohol have to do with decarbonisation?

“If the reactant molecules are rotated, then in principle, before and after this rotation, they can still undergo the same chemical reaction. But in the traditional machine-learning approach, the model will see these as two different reactions. That makes the machine-learning training much harder, as well as less accurate,” explains lead author Dr Chenru Duan.

The researchers aimed to build a program that could calculate transition states for two molecules in any orientation to each other. They used a generative AI tool called a diffusion model, and trained it on 9,000 chemical reactions. They fed the model the molecular shapes of the reagents, transition states, and products for each reaction.

“Once the model learns the underlying distribution of how these 3 structures coexist, we can give it new reactants and products, and it will try to generate a transition state structure that pairs with those reactants and products,” says Duan.

Duan and colleagues then gave the model 1,000 new reactions and told it to generate 40 possible transition states for each one. They used another model (a confidence model) to figure out which of these states were most likely.

The program was able to generate accurate transition states for each reaction in a few seconds.

“You can imagine that really scales to thinking about generating thousands of transition states in the time that it would normally take you to generate just a handful with the conventional method,” says Kulik.

The researchers used small molecules to train and test their reactions – up to 23 atoms for the whole reaction, which rules out most drug-making and anything involving a protein. But they found that the model could still work accurately for bigger molecules.

“Traditionally all of these calculations are performed with quantum chemistry, and now we’re able to replace the quantum chemistry part with this fast generative model,” says Duan.