💬 NLP

Catalyst-Agent: Autonomous heterogeneous catalyst screening and optimization with an LLM Agent

This paper introduces Catalyst-Agent, an LLM-powered autonomous agent that leverages Model Context Protocol and graph neural networks to efficiently screen, optimize, and propose novel heterogeneous catalysts for key reactions like ORR, NRR, and CO2RR with minimal human intervention.

Original authors: Achuth Chandrasekhar, Janghoon Ock, Amir Barati Farimani

Published 2026-03-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Achuth Chandrasekhar, Janghoon Ock, Amir Barati Farimani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find the perfect key to open a very specific, difficult lock. In the world of chemistry, this "lock" is a reaction we want to happen (like turning carbon dioxide into fuel or cleaning up pollution), and the "key" is a catalyst—a material that speeds up the reaction without being used up.

For decades, finding these keys has been a slow, expensive, and frustrating game of "guess and check." Scientists would mix chemicals in a lab, wait days for results, or run massive supercomputer simulations that took weeks. It was like trying to find a needle in a haystack by looking at one straw at a time.

Enter Catalyst-Agent, a new kind of AI robot scientist that changes the game entirely.

The AI Detective: Catalyst-Agent

Think of Catalyst-Agent not as a simple calculator, but as an autonomous detective with a superpower: it can talk to other specialized computer programs (tools) and use them to solve a mystery on its own.

Here is how it works, using a simple analogy:

1. The Brain (The LLM Agent)

At the center is a "Brain" powered by a Large Language Model (like a super-smart version of the AI you are talking to right now). But instead of just writing poems or answering trivia, this Brain has been trained to understand science.

The Job: You tell the Brain, "Find me a material that can turn Oxygen into water efficiently."
The Action: The Brain doesn't just guess. It formulates a plan, like a chess player thinking three moves ahead.

2. The Toolkit (The MCP Servers)

The Brain is connected to five specialized "limbs" or tools, each doing a specific job. Imagine a workshop where every tool is a different expert:

The Librarian: Goes to massive digital libraries of materials (like the Materials Project) to find candidate "keys" based on the Brain's request.
The Architect: Takes a raw material and builds a 3D model of its surface, like a construction worker setting up a stage.
The Tinkerer: If the first model isn't quite right, this tool tweaks it. It might swap one atom for another (like changing a screw on a machine) or stretch the material slightly (like stretching a rubber band).
The Tester: This is the heavy lifter. It uses a super-fast AI physics engine (called UMA) to simulate how the material behaves. It calculates the energy needed for the reaction in seconds, a task that would take a traditional supercomputer days.
The Judge: It looks at the test results. "Is this good enough? No? Okay, tell the Tinkerer to try again."

The "Closed Loop" Dance

The magic of Catalyst-Agent is that it does this in a closed loop. It doesn't need a human to say, "Okay, try this next."

Pick a candidate: The Brain picks a material from the library.
Test it: The Tester runs the simulation.
Evaluate: The Judge says, "It's close, but the energy is a tiny bit too high."
Iterate: The Brain thinks, "Okay, if I stretch the material by 2% and swap a surface atom, maybe it will work."
Repeat: The Architect and Tinkerer make the change, and the Tester runs it again.

This happens in a loop until the AI finds a winner or decides the material is a dead end. It's like a video game character that automatically adjusts its strategy when it hits a wall, rather than waiting for the player to tell it what to do.

The Results: Speed and Success

The researchers tested this AI on three major chemical challenges:

ORR: Turning oxygen into water (crucial for fuel cells).
NRR: Turning nitrogen into ammonia (for fertilizer).
CO2RR: Turning carbon dioxide into useful chemicals.

The Results were impressive:

Success Rate: About 1 in 3 materials the AI picked turned out to be good candidates. That's a huge win in a field where success rates are usually much lower.
Efficiency: On average, the AI only needed 1 to 2 tries to find a working solution for a successful material.
Discovery: It even found a new material (Sn3Sc) for turning CO2 into fuel that humans hadn't reported in the literature before!

Why This Matters

Before this, a human scientist might spend months setting up experiments or running one simulation at a time. Catalyst-Agent does the work of a whole team in a fraction of the time, for a very low cost (the paper notes the total cost for the AI's "thinking" was less than $70).

The Big Picture:
This isn't about replacing scientists. It's about giving them a super-powered assistant. Instead of spending their time writing code or waiting for computers to finish calculations, scientists can now just say, "Find me the best catalyst for X," and let the AI handle the heavy lifting, the trial-and-error, and the optimization.

It's the difference between manually digging for gold with a spoon and using a high-tech metal detector that tells you exactly where to dig. We are moving from an era of slow, manual discovery to an era of autonomous, fast-paced scientific invention.

1. Problem Statement

The discovery of novel heterogeneous catalysts is a critical bottleneck in chemical technology. Traditional approaches rely on:

Experimental Trial-and-Error: Time-consuming, expensive, and limited by physical synthesis capabilities.
First-Principles Computational Methods (DFT): While more efficient than wet labs, Density Functional Theory (DFT) calculations are computationally expensive (approx. 48 hours per relaxation) and struggle with the combinatorial explosion of material families, surface facets, and adsorption sites.
Current AI Limitations: Existing deep learning models (e.g., Graph Neural Networks) accelerate energy predictions but often lack the autonomy to orchestrate complex workflows. Previous LLM-based efforts (e.g., CatGPT, ChemReasoner) focused primarily on structure generation or literature mining rather than closed-loop, iterative optimization where an agent autonomously retrieves, evaluates, modifies, and re-evaluates materials based on quantitative feedback.

2. Methodology: Catalyst-Agent Framework

The authors introduce Catalyst-Agent, an autonomous AI agent built on the Model Context Protocol (MCP) framework. It integrates Large Language Model (LLM) reasoning with specialized computational tools to create a closed-loop screening pipeline.

A. Architecture

The system consists of a central LLM Supervisor (powered by GPT-5.2 via OpenAI's Codex CLI) acting as the orchestrator, connected to five specialized MCP Servers. This modular design isolates conflicting software dependencies and prevents GPU overloading.

Catalyst Information Server: Uses Perplexity Deep Research to generate literature-informed lists of candidate elemental compositions for specific reactions.
Structure Retrieval Server: Queries the OPTIMADE API (Materials Project, OQMD) to retrieve crystal structures and convert them into pymatgen/ASE objects.
CIF File Resource Server: Allows direct input of Crystallographic Information Files (CIFs) for custom or non-indexed materials, performing symmetry analysis.
Structure Modification Server: Enables the agent to apply targeted surface-level modifications to "near-miss" candidates:
- Top-layer substitution (Doping): Replacing a surface atom with a different element.
- In-plane strain: Uniformly scaling lattice vectors ( $a, b$ ) to induce compressive or tensile strain.
Energy Evaluation Server: The computational core. It uses FAIRchem's AdsorbML workflow with the UMA (Universal Machine-learned Interatomic Potential) small model (uma-s-1p1) to relax adsorbate-slab systems and calculate adsorption energies ( $\Delta E_{ads}$ ) for multiple facets in parallel.

B. Workflow

The agent operates in a closed-loop iterative cycle:

Proposal: Retrieve candidate materials based on user-defined reaction criteria.
Evaluation: Calculate adsorption energies for key intermediates on multiple facets.
Decision: Compare results against target descriptors (e.g., specific energy windows).
Refinement: If a candidate is a "near-miss" (close but not optimal), the agent autonomously hypothesizes and applies modifications (strain or doping).
Re-evaluation: Re-calculate energies for the modified structure.
Termination: Stop when criteria are met or the material is deemed unsuitable.

3. Key Contributions

Autonomous Closed-Loop Optimization: Unlike prior works that generate structures or mine literature, Catalyst-Agent performs end-to-end screening, including the iterative refinement of near-miss candidates via surface engineering.
MCP-Based Modularity: The use of the Model Context Protocol standardizes the interface between the LLM and heterogeneous computational tools (databases, ML potentials, simulation libraries), solving software dependency conflicts and enabling robust error handling.
Integration of ML Potentials: The workflow leverages UMA (Universal Machine-learned Interatomic Potentials) to approximate DFT accuracy at a fraction of the computational cost, enabling high-throughput screening.
Demonstration of "Scientific Reasoning": The agent demonstrates the ability to isolate variables (e.g., separating strain effects from doping effects) and adapt its strategy based on quantitative feedback, mimicking human experimental design.

4. Results

The agent was evaluated on three pivotal electrocatalytic reactions: Oxygen Reduction (ORR), Nitrogen Reduction (NRR), and CO2 Reduction (CO2RR).

Performance Metrics

Success Rate: The agent achieved a success rate of 23–34% across all evaluated materials (i.e., finding at least one condition satisfying the target descriptors).
- ORR: 34%
- NRR: 23%
- CO2RR: 31%
Efficiency: Successful materials converged in an average of 1–2 trials (evaluations) per material.
- ORR: 1.86 trials
- NRR: 2.08 trials
- CO2RR: 1.25 trials
Iterative Refinement: For materials requiring modification, the agent successfully identified productive modifications (doping/strain) more often than not for "near-miss" candidates. For example, in ORR, successful modified materials had a 1.67:1 ratio of successful to unsuccessful modifications, whereas failed materials had a 0.50:1 ratio.

Case Study: CoGa2O4 (ORR)

The agent successfully optimized the spinel CoGa2O4 for ORR.

Initial State: Baseline adsorption energy ( $E_{ads}$ ) was 1.11 eV (slightly outside the 0.9–1.1 eV optimal window).
Iteration 1: Attempted Ga→Al doping + 2% compressive strain. Result: $E_{ads}$ increased (worsened).
Iteration 2: Tested 2% compressive strain alone. Result: $E_{ads}$ increased further.
Iteration 3: Agent inferred the wrong direction and applied 2% tensile strain. Result: $E_{ads}$ dropped to 1.058 eV, successfully entering the optimal window.
Outcome: Convergence achieved in 4 trials, demonstrating adaptive reasoning.

Novel Discovery

The agent identified Sn3Sc as a potential novel CO2RR catalyst, a material not previously reported in the literature for this specific application.

5. Significance and Future Outlook

Accelerated Discovery: Catalyst-Agent reduces the time and human effort required for computational catalysis by automating the "hypothesis-test-revise" loop.
Scalability: The modular MCP architecture allows for easy integration of new tools (e.g., stability descriptors, microkinetic modeling) without rewriting the core agent logic.
Cost-Effectiveness: The study highlights that short-to-moderate screening campaigns are feasible at negligible cost (approx. $68 for the entire study using GPT-5.2), lowering the barrier to entry for AI-driven materials discovery.
Limitations & Future Work:
- Thermodynamic vs. Operating Conditions: Current descriptors rely on adsorption energies; future iterations need to incorporate solvent effects, applied potentials, and stability (phase segregation).
- Modification Space: Current modifications are limited to single-atom doping and uniform strain. Future work could explore complex defect engineering or multi-site substitutions.
- Hallucination & Reasoning: While the agent converges numerically, its verbal reasoning sometimes inverts thermodynamic logic (e.g., mislabeling weak binding as "too strong" but correcting via numerical tracking). Future systems may require tighter integration of physics-informed constraints to ensure reasoning consistency.

Conclusion: Catalyst-Agent represents a significant step toward fully autonomous scientific discovery, demonstrating that LLM agents can effectively orchestrate complex, tool-grounded workflows to identify and optimize catalyst materials with minimal human intervention.