In computational catalysis, adsorption energy is a fundamental descriptor used to assess and screen catalytic activity. However, because an adsorbate can bind to a surface in multiple configurations—varying by site and orientation—accurately determining the lowest (most stable) adsorption energy often requires evaluating a large number of possible configurations. Traditional approaches rely on exhaustive enumeration of these configurations, making the process both computationally expensive and inefficient, with no guarantee of locating the true global minimum. In this study, we present Adsorb-Agent a novel Large Language Model (LLM)-driven system designed to intelligently identify stable adsorption configurations with minimal computational effort. Instead of relying on brute-force sampling, the LLM leverages its general knowledge and reasoning capabilities to propose targeted configurations that are more likely to yield low adsorption energies. This strategy not only reduces the number of initial configurations needed but also improves the likelihood of identifying energetically favorable states. We benchmark Adsorb-Agent across twenty catalytic systems relevant to the oxygen reduction, nitrogen reduction, and hydrogen evolution reactions. The agent matches or improves upon the adsorption energies found by traditional methods in 83.7% of cases, and achieves lower (closer to global minimum) energies in 35% of systems tested. Notably, its advantages are most pronounced in complex scenarios: it outperforms standard approaches in 46.7% of systems featuring intermetallic surfaces and in 66.7% of systems involving large adsorbate molecules. These results highlight the promise of LLMs in accelerating catalyst discovery and solving intricate materials science problems—without the need for task-specific fine-tuning.
