An open dataset of gas adsorption in deformable metal-organic frameworks for direct air capture
The paper introduces the Open DAC 2025 (ODAC25) dataset, comprising nearly 60 million DFT calculations on approximately 15,000 deformable metal-organic frameworks to enable high-resolution machine learning-driven screening for direct air capture materials.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer
To understand the challenge of cleaning carbon dioxide from the air, one must first understand the materials scientists use to catch it. Imagine a sponge, but instead of being made of soft foam, it is built from metal atoms connected by organic molecules into a rigid, three-dimensional lattice. These are metal-organic frameworks, or MOFs. They are like microscopic sponges with pores so uniform and vast that they can trap specific gas molecules while letting others pass. For years, researchers have hoped to use these materials to pull carbon dioxide directly from the atmosphere, a process known as direct air capture, to help slow climate change. The problem is that there are millions of possible ways to build these sponges, and testing them one by one in a lab is too slow and expensive. Scientists have turned to computers to simulate how different MOFs behave, but for a long time, these simulations relied on simplified rules that often missed the subtle, real-world ways these materials bend and change when gas molecules enter them.
A new study introduces a massive, open collection of data designed to fix these gaps and accelerate the search for better air-cleaning materials. The researchers, a team spanning institutions from Meta to the Georgia Institute of Technology, have created a dataset called Open DAC 2025. This collection contains nearly 60 million detailed computer calculations describing how four different gases—carbon dioxide, water vapor, nitrogen, and oxygen—interact with about 15,000 different metal-organic frameworks. Unlike previous efforts that treated these materials as perfectly rigid cages, this new dataset accounts for the fact that the frameworks actually flex and shift when gas molecules enter them. By including these movements, along with new types of chemical modifications and even computer-generated structures that do not yet exist in nature, the team has provided a more realistic and comprehensive map for finding the best materials to capture carbon.
The core of this work addresses a specific flaw in how scientists previously modeled these materials. In earlier studies, researchers often assumed the metal-organic framework stayed exactly the same shape whether it was empty or filled with gas. In reality, when a gas molecule enters a pore, it can push the framework's atoms slightly, causing the material to deform. This deformation changes the energy required to hold the gas. The new dataset corrects this by running simulations where the framework is allowed to relax and find its most stable shape after a gas molecule enters, and then again after the gas leaves. This process revealed that many materials previously thought to be strong candidates for carbon capture might actually be less effective once their flexibility is taken into account. The researchers found that ignoring this movement could lead to significant errors in predicting how well a material works, sometimes making a material look much better than it truly is.
To build this dataset, the team did not just look at existing materials; they expanded the chemical space in several ways. They added nitrogen and oxygen to the mix, gases that are abundant in the air and compete with carbon dioxide for space inside the pores. They also created new versions of the frameworks by attaching amine groups—chemical structures known to attract carbon dioxide—to the metal sites or the connecting links of the framework. Furthermore, they used computer algorithms to generate thousands of hypothetical structures that are chemically valid but have not yet been synthesized in a lab. These synthetic structures explored geometries with larger pores and lower densities than what is currently found in experimental databases, effectively testing the limits of what these materials could theoretically be.
The sheer volume of data required to train the next generation of artificial intelligence models is staggering. The team performed nearly 60 million individual calculations, a task that would be impossible to complete in a lab setting. They used a method called density functional theory, which is a high-accuracy way of calculating how electrons behave in atoms and molecules, to determine the energy of every interaction. To ensure these calculations were precise, they corrected for numerical errors that can occur when simulating the repeating patterns of the crystal structures. They also included high-energy, unstable configurations of gas molecules inside the pores, which are necessary for training computer models to predict how the materials behave under a wide range of conditions, not just in their most comfortable state.
With this dataset in hand, the researchers trained several advanced machine-learning models to predict the properties of these materials without needing to run the expensive calculations every time. These models, which act as a shortcut for the complex physics, were tested on their ability to predict how much gas a material would hold at low pressures, a key measure of performance. The results showed that models trained on this new, more realistic data were significantly more accurate than those trained on older, simplified data. They were better at predicting the behavior of the materials, including how the frameworks deformed and how the different gases competed with one another. This improvement is crucial because it means scientists can now screen thousands of potential materials with a level of confidence that was previously unattainable.
The study also highlighted the importance of including all the gases found in the air, not just carbon dioxide. By simulating how nitrogen and oxygen interact with the frameworks, the researchers could identify materials that might be selective for carbon dioxide even in the presence of these other gases. They found that some materials showed a strong preference for nitrogen or oxygen, which would be a problem for carbon capture applications. This level of detail helps researchers avoid wasting time on materials that might look good in a simple test but fail in the complex environment of real air.
While the dataset is a powerful tool, the researchers are careful to note its limitations. The simulations rely on specific mathematical approximations for how electrons interact, which can sometimes struggle with certain types of chemical reactions, particularly those involving oxygen. They acknowledge that for the most critical applications, further testing with more complex methods will be needed. However, the dataset provides a robust foundation that is far superior to the simplified models used in the past. By making this data and the trained models publicly available, the team hopes to empower other scientists to design better materials faster, moving the field closer to practical, large-scale solutions for removing carbon dioxide from the atmosphere.
The work represents a shift from looking at these materials as static objects to understanding them as dynamic systems that change in response to their environment. This shift allows for a more honest assessment of which materials will work in the real world. The dataset includes nearly 60 million data points, covering a vast array of chemical possibilities, from the materials we know today to those we have only imagined. It serves as a comprehensive guide for the next generation of discovery, ensuring that when new materials are built, they are built on a foundation of accurate, detailed, and realistic science. The ultimate goal is to find a material that can efficiently and cheaply pull carbon dioxide from the sky, and this new resource provides the clearest map yet to find it.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.