Science

Meta’s mega gift to material science – BusinessLine


Discovery of new materials with desired properties is not easy, ask any alchemist. You must take two materials that have the properties that you want and, like making a baby, you must see how to combine them to get exactly the offspring that you want. 

Meta (formerly, Facebook), often bashed for cashing-in on data, has now given a meta gift to mankind: tomes of data — for research into functional (meta) materials. 

The global giant has just released copious amounts of data about behaviour of materials at the atomic level, under its Open Materials 2024 (OMat24) initiative, for free. 

This data can help scientists combine different materials with different desired properties to create something new. Examples of functional materials that can tackle climate change include new catalysts for renewable energy storage, carbon neutral fuels, new sorbents for direct air capture, etc. 

Creating such materials is somewhat similar to discovering a drug molecule for disease by trying out different combinations of molecules or making a dish of desired taste, texture and flavour from millions of ingredients. 

Traditionally, making meta materials involved playing trial-and-error with millions of data points from thousands of materials. But now, there is Artificial Intelligence (AI), which can deliver the goods, however, AI is data hungry.  

Meta has provided this food for AI. The OMat24 dataset is a collection of data generated from simulations and calculations on different inorganic materials. This dataset contains information on 118 million atomic structures, whichhas information on three parameters — total energy (the overall energy of the material’s structure), forces (acting on each atom) and cell stress (indicating how the material could deform under certain conditions.) All this was calculated using ‘density functional theory’, a quantum mechanics-based method for predicting material properties . 

“The search space of possible materials is enormous and remains a significant challenge for both computational and experimental approaches to material science,” says a yet-to-be peer-reviewed paper by scientists at Meta’s Fundamental AI Research (FAIR) group. “Identifying promising candidates through computational screening with machine learning models offers the potential to dramatically increase the search space and the rate of experimental discovery,” says the paper, written by Luis Borroso-Luque et al. 

FAIR scientists took over 400 million core-hours of computing to get the data of 118 million structures labelled with total energy, forces and cell stress. (One CPU core used for one hour is one core-hour.) These parameters give an idea of the stability of materials under given conditions. 

Now that there is data on over 118 million atomic structures, one can train an AI model to come up with the best combination of any of them for a desired material. 





READ SOURCE

This website uses cookies. By continuing to use this site, you accept our use of cookies.