Skip to main content
Computing, Environment and Life Sciences

Applying Machine Learning to Understand Water’s Weird” Properties

Argonne and Pacific Northwest national laboratories are applying a suite of machine-learning techniques to better understand why water behaves so much differently than other liquids.

The intricate structures that water forms at the atomic scale make it a strange liquid. Each water molecule has two pairs of sites – donors and acceptors – that, like poles on a magnet, link it to other water molecules. The odd shape of water molecules and these four linking sites cause water to arrange itself in complex networks that result in unique properties such as a high boiling point. Besides explaining water’s behaviors, understanding how intermolecular and long-range interactions build such networks is central to phenomena as diverse as gene regulation, topological states of quantum materials, electrolyte transport in batteries, and – of course – the universal solvation properties of water.

A team from Argonne and Pacific Northwest national laboratories is working to help understand these networks using a suite of machine learning tools. As highlighted in a recent paper at NeurIPS, predicting how networks form in water requires researchers to solve gradually harder challenges, including how to predict the stability of a network if you know the location of each atom and they are bonded. The grand challenge is generating realistic water cluster networks from scratch.

The paper describes these challenges and presents a suite of initial solutions for the larger AI community, along with a recently published dataset of 4.95 million water clusters. The team hopes that the open dataset and clear, impactful definition of challenges will drive new AI research and provide benefits for the chemistry community.

Determining the stability of water molecules is an enticing problem that requires researchers to learn the effects of both short- and long-distance interactions simultaneously. Generating feasible structures from scratch could also require advances in optimization, reinforcement of learning algorithms, and determination of how to deploy the algorithms on next-generation computers. If solved, a thorough understanding of these complex networks can be used to design better solvents and be applied to a range of science fields, like designing better electrical grids. Taken together, DOE’s water and AI research are a great mix.