• In collaboration with our partners at the University of Toronto and VSParticle, Meta FAIR is releasing a large-scale dataset of experimental results on various materials, providing valuable insights for the development of new catalysts.
• This work is a continuation of Meta FAIR’s Open Catalyst Project which seeks to use AI to model and discover new catalysts to address the energy challenges posed by climate change.

Understanding how atoms interact with each other could lead to numerous scientific advances, including converting waste CO2 into valuable chemicals, enabling the creation of sustainable aviation fuels, new pharmaceuticals, consumer products from renewables, and clean hydrogen for various applications. The challenge is that simulating sets of atoms is computationally very expensive, especially when modeling large numbers of atoms or how atomic systems evolve over longer time scales. To address this and push research in the field forward, over the last five years, Meta FAIR’s chemistry team has released multiple datasets and AI models aimed at significantly speeding up these calculations with our Meta Open Catalyst, Meta Open DAC, and Meta Open Materials efforts. Unlike other industry-led efforts, which are often proprietary, these releases bring comprehensive datasets and state-of-the-art models to the broader academic community, allowing for more accurate and scalable research.

We’re now taking a crucial next step in the lab, translating theoretical concepts into practical applications. Computational models generally provide simplified representations of physical-world scenarios, but their results still need to be proven in actual laboratory settings. Building on complementary expertise and resources, we partnered with the University of Toronto and VSParticle, a Dutch nanotechnology engineering company, to establish a streamlined pipeline for synthesizing and testing a large number of mixed metal catalyst materials. This two-year effort resulted in a dataset of over 600 materials to help the community start to bridge the gap between computational predictions and experimental results.

Experimentally demonstrating the effectiveness of computational predictions is a complex process. It requires the synthesis of a wide variety of catalysts, confirmation that the materials generated match the desired target, and finally the testing of the catalysts in industrially relevant conditions. We synthesized catalysts containing 13 different elements and verified their composition and structure using various characterization techniques. This process is highly challenging, with less than 25% of the catalysts made matching the desired targets. To expand the diversity of the materials, two different automated synthesis techniques were used. The first technique—enabled by VSParticle—vaporizes metal rods to create nanoparticles of the desired composition. Their VSP-P1 printer provides high levels of automation and speed to create large quantities of new materials. The second approach uses an automated robotic system developed by the University of Toronto to perform chemical reduction through wet chemistry.

For testing, the University of Toronto developed a high-throughput testing pipeline, which enabled 30 experiments to be performed per day. We tested the catalysts for green hydrogen production with the goal of identifying new, low-cost catalysts that can replace expensive platinum-based catalysts for Hydrogen Evolution Reaction (HER). We also tested for CO2 recycling using CO2 Reduction Reactions (CO2RR) with the goal of finding catalysts that are selective to specific and useful products.

For HER, our models predicted platinum and Pt-containing alloys, which are known to be effective but expensive catalysts, as promising candidates even though no platinum was contained in any of the training set’s materials. Through this analysis of over 19,000 materials, we identified hundreds of potential HER catalysts, many of which are importantly composed of low-cost elements. This approach has the potential to significantly lower the cost of hydrogen production, increasing its viability as a clean energy source.

For CO2RR, we tested the catalysts at current densities that are relevant for industrial conditions to gain a better understanding for how they might perform in practical settings. The predictions for this more complex reaction were less correlated with the experimental results. However, the dataset offers an opportunity for the community to further improve and benchmark computational models.

An ideal experimental dataset for training AI models should be reproducible with positive and negative results over a diverse set of catalyst materials.The common practice of only releasing positive results and performing experiments under different conditions makes obtaining such a dataset difficult from current literature. We are directly addressing that lack of data by releasing our full experimental dataset with a wide variety of catalysts tested under the same conditions with the goal of driving progress in this important field. We view our work as a pilot demonstration of what’s possible and hope to continue to build larger and more diverse experimental datasets to help solve community challenges.

In addition to our experimental data, we’re also releasing the computational analysis of over 19,000 catalyst materials. This includes how strongly the catalysts attract various molecules that are involved in our reactions of interest. While AI has greatly reduced the amount of time required to perform these calculations, the sheer magnitude of the number of calculations (over 685 million relaxations) required would still be beyond the means of most labs. Due to this, we’ve pre-computed these values and are releasing them publicly for the community to build upon.

From the beginning of our chemistry and material science efforts at Meta with the Open Catalyst project, we’ve focused on climate applications such as renewable energy storage, green hydrogen generation, and the generation of fuels from renewable energy. Finding low-cost and effective catalysts for the reactions that drive these processes is essential for enabling a carbon net-zero future. Success in this research challenge offers the opportunity for world-changing impact, similar to the study of proteins for drug discovery.

Looking forward, we’re excited to continue our work leveraging AI that may help mitigate the effects of climate change, as well as continuing to explore other impactful applications of AI in material science discovery. The potential applications are wide-ranging. For example, as well as this work on catalysts, we’re currently collaborating with Meta’s Reality Lab Research to model materials at the atomic level to discover novel crystals that can be used to drive innovation in AR glasses. We strongly believe that, as the AI models become more efficient and improve in their ability to generalize across a wide variety of materials and molecules, we’ll see breakthroughs across various industries that will have a profound impact on humanity and how we interact with technology.