Neel Redkar—Engineering Project Template
Summary
Carbon capture is key to reversing climate change by trapping CO2 emissions from factories. Electrocatalysis of CO2 is a new approach in which the trapped CO2 is split into usable CO and O2. The current solutions are costly and inefficient. Metal Organic Frameworks (MOFs) are crystals made of organo-metals and are adept at adsorbing gases and catalyzing reactions. The engineering goal for this project was to build an active learning neural network using available sparse data to design a MOF that can efficiently adsorb CO2 and convert it to CO and O while maintaining a low manufacturing cost. Iteratively, a Se-MOF (C₁₈MgO₂₅Se₁₁Sn₂₀Zn₅) was converged upon, shown to be more effective and less costly than existing MOFs, as well as supported by the current literature. This novel network could be implemented for other gas separation and catalysis applications using a sparse dataset.

Introduction
Background
The Problem
- 51% of total US carbon emissions are from factories in the form of flue gas that is high in carbon dioxide (CO2) concentration (Rahimi et al.)
- Capturing CO2 at source is critical to halting climate change. The by-product, clean filtered steam, can be reused to decrease net energy use by 25-40%. (Rahimi et al.)
- Two technologies are used for CO2 capture - absorption using liquid amines and adsorption on solids. Both technologies have shortfalls and active research to find the best solution is ongoing.
- Liquid amines are toxic to the environment and are corrosive; they can be vaporized and need frequent replacement. (Ünveren et al.)
Current Research on Solid Adsorbents
- Metal Organic Frameworks (MOFs) are crystals made of organo metals to adsorb and filter carbon dioxide in solid state (no corrosion or vaporization)
- MOFs can have ligands to catalyze reactions and break down CO2 with electrocatalysis (by-products are CO, CH4) (Shao et al.)
- Potential to lower CO2 emissions, improve capture efficiency & make usable by-products, thus a practical and sustainable option
- Empirical data on existing MOFs is sparse, so researchers are trying to use generative neural networks to build novel MOFs
- Cost is not often included into existing prediction studies
Active-Transfer Learning vs Current Models
- Most popular models utilized for generation are AutoEncoders and Generative Adversarial Neural Networks (Chong et al.)
- The downside to this is that they need large amounts of data to learn the design space (Kim J et al.)