With climate change accelerating, renewable energy sources like wind and solar are vital to the modern energy grid. Shifting to renewable energy requires a way to store power for the times when the sun isn’t shining and the wind isn’t blowing. Doing this requires electrocatalysts. However, the ones available today are inefficient or rely on rare and expensive materials.
Open Catalyst aims to discover low cost catalysts that can drive the chemical reactions necessary to convert excess solar and wind energy into other fuels that may be used to generate electricity when other sources of renewable energy are unavailable. By developing AI to accurately predict atomic interactions faster than the compute-heavy simulations scientists rely on today, calculations that take modern laboratories days could instead take seconds. This would enable the large-scale exploration of new materials which is crucial when there are billions of possible combinations to test.
I'm a very outdoorsy person, so seeing the environment and climate change, seeing the smoke and fires in California, to me feels very personal. My daughter and I talk about this and she gets stressed out because it's her that's going to have to deal with a lot of these issues.
Environmental and climate issues are increasing all around the world. I think it’s important to invest in AI related solutions to address some of these challenges as best as we can. I’m really glad I can help investigate this avenue of work.
A few years back when I was looking for a stronger lever for real world impact, one of the problem spaces that seemed urgent was climate change. So I asked myself where I could make an impact in addressing climate change with my expertise in machine learning.
Our researchers see the impact of climate change every single day. Whether it be unrelenting fires in California or air pollution caused by fireworks in India, they want to combat the negative environmental impact not only for themselves and their children, but for everyone in the world. With their expertise in machine learning, they’re striving to find new ways to provide a scalable alternative to expensive storage technologies so that clean and sustainable power can be supplied the world over. By sharing the Open Catalyst 2020 (OC20) dataset with the scientific community, they hope it will lead to accelerated progress on this critical undertaking.
Finding the right combination of catalysts is a time-consuming process. There are billions of possible combinations of elements to try. Experimentalists using standard synthesis methods can try 10 materials per day, while a modern computational laboratory using quantum mechanical simulation tools such as density functional theory (DFT) can run 40,000 simulations per year. The goal of Open Catalyst is to enable researchers to screen millions and maybe even billions of possible catalysts per year.
The Open Catalyst team uses a comparatively small number of DFT calculations to train more efficient ML models on the fundamental physics governing quantum mechanics to learn how to predict the forces exerted on atoms by each other and the overall system energy. Focusing on materials and molecules that are important in renewable energy applications, the OC20 dataset is the largest dataset of electrocatalyst structures to date and as such should lead to significant improvements in ML models, specifically in their ability to generalize and learn the underlying physics governing molecules at inorganic interfaces. It also opens the door to predicting reaction selectivity across catalyst composition.
C. Lawrence Zitnick, Lowik Chanussot, Abhishek Das, Siddharth Goyal, Javier Heras-Domingo, Caleb Ho, Weihua Hu, Johannes Klicpera, Adeesh Kolluru, Janice Lan, Thibaut Lavril, Aini Palizhati, Morgane Riviere, Ammar Rizvi, Alex Schneidman, Nima Shoghi, Muhammed Shuaibi, Anuroop Sriram, Kevin Tran, Richard Tran, Brook Wander, Brandon Wood, Junwoong Yoon, Jure Zbontar, Devi Parikh, Zachary Ulissi
Through our vision and research in ML, we developed wav2vec, a way to build speech recognition systems that require no transcribed data.
Meta AI and NYU Langone Health have developed a way to use AI for MRIs that need only a quarter of the raw data traditionally required for a full MRI.