From Data to Physics: An Agentic Large Language Model Solves a Competitive Adsorption Puzzle

Scientific modeling often requires navigating a trade-off between physical interpretability and empirical accuracy-a task that can take weeks of iteration, especially in systems with partial observability, structural complexity, and experimental errors. Here, we show how an agentic reasoning-and-coding large language model (LLM), OpenAI o3, autonomously solved a modeling challenge in surface chemistry that puzzled us for months: quantifying the competitive adsorption of carboxylic acids on metal-organic layers (MOLs). With experimental data and a concise problem formulation, o3 rapidly formulated a physically grounded adsorption model, derived the mathematical equations, implemented the corresponding codes to fit the experimental data, revised its assumptions, and ultimately derived a competitive adsorption model with three parameters that matched experimental data across more than a dozen tested molecules. The resulting model-simple, mechanistically transparent, and quantitatively robust-incorporates both classical Langmuir competition and structural constraints such as site accessibility. Beyond addressing this particular challenge, our findings highlight a transformative shift in scientific methodology: from manual trial-and-error approaches to AI-driven hypothesis generation and model refinement. This represents a new paradigm in research, wherein language models surpass the traditional roles of machine learning in data analysis and computational support, actively participating in scientific reasoning and hypothesis development.