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

Area of Science:

Surface Chemistry
Computational Chemistry
Materials Science

Background:

Scientific modeling often involves a difficult trade-off between physical interpretability and empirical accuracy.
Complex systems with partial observability, structural complexity, and experimental errors exacerbate this challenge, requiring extensive manual iteration.
Quantifying competitive adsorption in such systems, like carboxylic acids on metal-organic layers (MOLs), has been a long-standing problem.

Purpose of the Study:

To demonstrate the capability of an agentic reasoning-and-coding large language model (LLM) to autonomously solve complex scientific modeling challenges.
To develop a physically grounded and empirically accurate model for competitive adsorption of carboxylic acids on MOLs.
To showcase a new paradigm in scientific research driven by AI hypothesis generation and model refinement.

Main Methods:

Utilized an agentic reasoning-and-coding LLM (OpenAI o3) with experimental data and a problem formulation.
The LLM autonomously formulated a physically grounded adsorption model, derived equations, and implemented fitting codes.
Iterative refinement of assumptions by the LLM led to the final competitive adsorption model.

Main Results:

The LLM successfully quantified the competitive adsorption of carboxylic acids on MOLs, a problem that had puzzled researchers for months.
A simple, mechanistically transparent, and quantitatively robust three-parameter model was derived.
The model accurately matched experimental data across more than a dozen tested molecules, incorporating Langmuir competition and structural constraints.

Conclusions:

Agentic LLMs can autonomously solve complex scientific modeling challenges, significantly accelerating research.
This represents a transformative shift from manual trial-and-error to AI-driven hypothesis generation and model refinement in scientific methodology.
LLMs are emerging as active participants in scientific reasoning, moving beyond traditional data analysis and computational support roles.