Reducing Carbon Footprint with AI

In this article, I will introduce the concept of direct air capture of CO2 (Carbon Dioxide) using a special class of new materials called MOFs. These materials can store certain gases via adsorption. This class of materials can be synthetically generated by choosing from a large combination of metals and organic compounds. Here, AI algorithms can be applicable in two ways:

1. Generation of stable versions of these structures.
2. Prediction of the efficiency at which these structures can adsorb (store) carbon dioxide under different conditions.

Broadly, we will look at three aspects:

1. What are climate change and direct air capture?
2. What kind of role can AI play in materials discovery, and are there some informative resources on these?
3. A dataset that can be used to build models for predicting the properties of MOFs for the direct air capture of CO2.

Lets start the very beginning — what is climate change?

In one sentence, climate change is the increase of the average temperature across the planet. Studies have shown that the global temperature is 1.1 degrees higher now than it was 100 years ago. It is this change in temperature that has been causing extreme weather, heat waves, melting glaciers, which in turn is leading to sea level rise. This rise in global temperature exactly coincides with our increased energy usage, most of which has come from fossil fuels. Fossil fuels release gases into the atmosphere such as — CO2, CO, and CH4. An excess of greenhouse gases traps heat in the earth, heat that would have otherwise dissipated into the atmosphere. The atmospheric carbon dioxide is not only rising, but it is rising at a faster rate, accelerating the amount of heat that is getting trapped. Information in the form of charts explaining the global temperature rise, carbon dioxide concentration in the atmosphere, and rise in sea temperature is publicly available. [1]

The challenge here is that carbon dioxide is hard to get rid of. After a pulse of CO2 is emitted into the atmosphere, 40% will remain for 100 years, 20% will reside for 1000 years, and the final 10% will take 10,000 years to turn over. [2]

Ok, so how do we get rid of all this CO2?

Action has been taken by several governments and individuals to slow down the temperature increase, the goal being to keep it to 1.5 degrees Celsius, as defined in the Paris Agreement. This involves increased focus on energy conservation, increasing investment in alternative sources of energy including renewable energy systems such as solar and wind, a radical shift to electric vehicles, and a focus on ESG declarations by public companies.

However, as discussed, CO2 stays in the atmosphere for a long time; hence, we may need tactical solutions as well. This is where carbon capture comes in.

Carbon capture involves capturing CO2 from the source of emissions and storing it permanently (sometimes under the surface of the earth). It is a controversial solution — as it does not solve the core problem. Direct air capture, involves separating carbon dioxide from ambient air using solid and liquid sorbents. This is an emerging area, as it involves capturing low concentrations of carbon dioxide from ambient air, and requires the discovery of materials that can capture CO2 at room temperature and at various humidity levels.

MOFs: What are they, and why are they so special?

Metal-Organic Frameworks (MOFs) are composed of two main types of components:

1. Metal Ions or Clusters: These are the inorganic part of MOFs. Metal ions or clusters act as joints or connectors in the framework. Different types of metals can be used, which affects the properties of the MOF. Common metals used include zinc, copper, iron, and aluminum.
2. Organic Linkers: These are organic molecules (containing carbon) that act like the rods or bridges between the metal ions or clusters. These linkers are usually composed of carbon, hydrogen, oxygen, and sometimes nitrogen. The type and arrangement of these organic molecules can be varied, which allows for a lot of diversity in MOF structures and properties.

When these metal ions or clusters are connected by the organic linkers, they form a three-dimensional structure with a lot of open space, like a sponge. This structure is what gives MOFs their high porosity and large surface area. The ability to mix and match different metals and organic linkers allows scientists to design MOFs for specific purposes, like gas storage, catalysis, or drug delivery.

Hence, MOFs are highly tunable and can capture gases and release them while using a very low amount of energy.

Broadly, where does AI fit in?

AI is making its way into the chemistry and biosciences domains — and shows potential for accelerating new scientific discoveries. Recently, researchers at MIT made headlines when they used neural networks to discover a new drug that could work against antibiotic-resistant bacteria. Interestingly, this drug had a chemical structure different from that of existing antibiotics. [4]

To bring the power of AI in materials discovery into perspective, a project released by DeepMind used an AI active learning-based pipeline (first of its kind) to find 380,000 new stable materials (from a total of 2.2 million predictions) that can be tested experimentally. As a comparison — existing databases of experimentally verified and computationally verified inorganic crystals totaled only 68,000 (20,000 + 48,000). These inorganic materials could be used for many applications, from powering supercomputers to next-generation batteries. [3]

if we consider MOFs specifically, AI can be used to:

1. Query certain properties of MOFs, including details of their structure and synthesis methods, which may be present in structured databases or unstructured form like published papers. [5] [6]
2. Generate synthetic versions of the MOFs from text specifications. [7]
3. Predict certain properties of MOFs (like their adsorption capacity) based on the structural details.[8]

How can I identify MOFs for Direct Air Capture?

Traditional computational studies of materials use Density Functional Theory calculations (DFT). DFT offers comprehensive insights into the electronic structure and atomic-level bonding of materials. This insight helps explain why a material displays specific properties and helps in forecasting how modifications to the material’s structure may influence its behavior. Experimentally assessing each potential new material is time-consuming and costly. DFT helps narrow down the extensive pool of potential materials to the most promising candidates, which can undergo synthesis and experimental testing.

Although DFT has helped speed up materials discovery, it can still take anywhere from days to weeks to discover the required properties of a single material and involves very specific expertise to use the DFT-related frameworks and software. ML/AI approaches can help speed this up by using trained graph neural networks to represent the material and predict the properties of the material relevant to the application area. Graphs are one efficient representation of molecules or materials, but several others exist.

The reason we can consider machine learning methods is because we now have huge amount of DFT calculations of materials of various types — including MOFs [9]. Open DAC 2023 (ODAC23) is a substantial dataset comprising over 38 million Density Functional Theory (DFT) calculations performed on a collection of over 8,400 MOF materials with adsorbed CO2 and/or H2O. It stands out as the most extensive dataset of MOF adsorption calculations using DFT. Noteworthy here is the fact that databases exist which contain DFT calculations of MOFs alone and DFT calulations of MOFs with CO2 adsorption, but this is the first one which takes into account adsorption of CO2 on MOFs in the presence of water. [10]

So, how can MOFs reduce our carbon footprint?

As discussed, MOFs are a unique class of compounds that are amenable to customization and can be designed to store gases through a process called adsorption. We need to reduce the CO2 in the atmosphere to slow down the rise of Earth’s temperature. Designing the most efficient MOF for storing carbon dioxide can be achieved using ML/AI models trained on a large number of existing MOF structures.

References:

[1] https://climate.metoffice.cloud/dashboard.html.

[2] https://www.ucsusa.org/resources/why-does-co2-get-more-attention-other-gases

[3] https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/

[4] https://news.mit.edu/2020/artificial-intelligence-identifies-new-antibiotic-0220

[5] https://github.com/Andrew-S-Rosen/QMOF

[6] DigiMOF: A Database of Metal–Organic Framework Synthesis Information Generated via Text Mining: https://pubs.acs.org/doi/10.1021/acs.chemmater.3c00788

[7] ChatMOF: An Autonomous AI System for Predicting and Generating Metal-Organic Frameworks: https://github.com/Yeonghun1675/ChatMOF

[8] A multi-modal pre-training transformer for universal transfer learning in metal–organic frameworks: https://www.nature.com/articles/s42256-023-00628-2

[9] https://next-gen.materialsproject.org/

[10] The Open DAC 2023 Dataset and Challenges for Sorbent Discovery in Direct Air Capture: https://arxiv.org/pdf/2311.00341.pdf#:~:text=Metal%2Dorganic%20frameworks%20(MOFs),functions%20of%20humidity%20and%20temperature.