Technologies designed to prevent more CO2 from entering the atmosphere must be as efficient as possible to not create new problems for our climate and environment.
The most widely used CO2 capture technology is absorption, where flue gas containing CO2 is absorbed into a solvent that consists of various chemical components. The basic solvent dissolves the acidic CO2, trapping it in the liquid. The purified flue gas is released into the air, while the CO2-rich solvent is heated above boiling point to release the CO2, which can be stored or utilized.
At DTU, Professor Philip Fosbøl is leading an EUDP-funded project that uses artificial intelligence, AI, to find new solvents that can handle this process more efficiently and thus more cheaply than current solutions.
"Our existing solvents are not optimal. It requires a lot of energy to extract the CO2 from the liquid so that it can be prepared for reuse. That is why we are working to find alternative molecules that can be used in an improved solvent," he explains.
AI does the heavy lifting
The project focuses on solvents made from amines a class of substances that always contain molecules with nitrogen at their centre. However, individual molecules can be constructed in a myriad of ways by replacing some of the atoms, just as the individual molecules can be combined in an enormous number of ways, akin to building with a pile of Lego bricks, where the bricks can be put together in an almost infinite number of ways.
"The challenge is the vastness of combination possibilities when putting together a solvent. Research has previously reached its limit because the only way to go about it has been to enter the lab and conduct oceans of experiments with possible candidates one by one. Instead, we are trying to implement a kind of synthetic test method using an AI," Philip Fosbøl explains.
The AI which has been specially developed for this purpose by project partner Hafnium Labs has been trained on published information on solvents and data from computer simulations so that it can identify the most promising candidates.
It does so by making a series of calculations on data and knowledge-based assumptions about the properties of both known and simulated molecules, such as melting point and heat of vaporization, as well as how the different molecules affect each other.
A needle in a massive haystack
The aim of the project is to screen more than a billion possible solvents and the AI has already suggested the first approx. 100,000 candidates. Researchers at DTU have carefully examined these to select two that are now being tested in the lab.
In this phase, the researchers must determine which components are included in the solvent and in what concentrations, as the AI does not provide the recipe for the solvents, explains researcher Randi Neerup, who has daily oversight of the project.
She says that if the work in the laboratory shows promising results, the researchers will move into DTU Chemical Engineering's pilot plant to test the solution on a larger scale and validate its usability before it can be tested at project partner Ørsted's facility.
Self-reinforcing process
Randi Neerup expects the testing phase for an alternative solvent to run for about a year, and that during the project there will be resources to test a total of ten candidates. Throughout the project period, work will also be done to optimize the AI, she explains:
"Our observations in the laboratory will be entered into the AI's database as new knowledge and then, in a year's time, with the new knowledge we feed in, we may get new suggestions that we should examine further."
The project partners expect that the work on further developing the AI during the project will also enable Hafnium Labs to make it applicable not only for CO2 capture, but also for other types of molecular development within, e.g., the manufacture of cleaning products or foodstuffs.
