Predicting CO2 Adsorption Capacity in Amine-Functionalized Zeolites Using Bayesian Optimization with Language-Interfaced Fine-Tuning (BO-LIFT)

Amine-functionalized materials are promising adsorbents for the direct air capture of carbon dioxide (CO₂). These materials consist of porous, inorganic supports, such as silica or zeolites, that are functionalized with organic amines. The amine functional group acts as a base to provide binding sites during interactions with the acidic CO₂ molecule, while the inorganic scaffolds upon which they are immobilized provide high surface area and stability. There exists a gap in the literature in understanding the importance of synthesis variables such as loading, solvent, and support porosity on CO₂ adsorption capacity. To explore these relationships, we are leveraging large-language models (LLMs) to develop a method of representing the amine-based materials by the text of the associated synthesis procedure. By using Bayesian optimization with language-interfaced fine-tuning (”BO-LIFT”), we will make predictions on adsorbent performance with text. Synthesis and adsorption experiments are performed ad-hoc upon request from the LLM, and predictions are made via Bayesian optimization of the black-box function mapping synthesis procedures to adsorption capacity. Our approach to representing adsorbents with the text of synthesis procedures is verifiable and actionable; it simultaneously eliminates the need for structural data of complex adsorbents when making predictions. Through eight iterative testing loops from two initial points, we achieve a maximum prediction accuracy of 96.9%, revealing the potential for the BO-LIFT approach in improving adsorbent performance through the interface of natural language.