ChatMOSP: A Chemistry-Grounded Mobile Agent for Working-State Catalyst Simulations

This paper introduces ChatMOSP, a chemistry-grounded mobile agent that translates natural language requests into validated multiscale simulations of working-state catalysts by dynamically mapping reaction conditions to morphology and activity models, retrieving necessary parameters from databases or literature, and successfully replicating complex experimental phenomena like temperature-induced morphological transitions and oscillatory reaction behaviors.

The Gist

Imagine you have a tiny, magical chef (a catalyst nanoparticle) that cooks chemical reactions. This chef is very picky: their shape, mood, and cooking speed change depending on the heat, the pressure, and the ingredients (gases) in the kitchen. If the kitchen gets too hot or fills with too much of a specific gas, the chef might change from a sharp, angular cube into a smooth, round ball. This shape-shifting determines how well they cook.

For a long time, figuring out exactly how this chef behaves required a team of highly specialized experts who spoke a very difficult "computer language" to run complex simulations. If you wanted to know what the chef looks like at a specific temperature, you had to be a master programmer to set it up.

Enter ChatMOSP: The "Translator" Chef's Assistant

This paper introduces ChatMOSP, a new tool that acts like a super-smart, chemistry-savvy personal assistant. You don't need to be a computer expert to use it. You can just talk to it on your phone, either by typing or speaking, in plain English (or Chinese).

Here is how it works, using simple analogies:

1. The "Magic Translator"

Think of ChatMOSP as a translator standing between you and a complex machine.

• You say: "Show me what a Palladium (Pd) nanoparticle looks like when it's hot and surrounded by CO gas."
• ChatMOSP hears: It understands your everyday words and instantly translates them into the strict, mathematical instructions the simulation machine needs. It doesn't just guess the answer; it sends the correct commands to a powerful physics engine called MOSP (Multi-scale Operando Simulation Package) to do the real work.

2. The "Smart Librarian"

Sometimes, the machine needs specific numbers (like how strongly a gas sticks to the metal) that aren't already in its internal memory.

• The Problem: The machine says, "I don't have the data for this specific gas."
• The Solution: ChatMOSP acts like a super-fast librarian. It goes out to the internet, finds scientific papers (like searching a library catalog), reads the abstracts, and pulls the exact numbers it needs from those papers. It then double-checks them with you before using them. This means you can simulate new scenarios even if the data wasn't pre-loaded.

3. The "Mobile Lab"

The most exciting part is that this whole process happens on a mobile phone. You can walk around, ask questions, and get answers about how these tiny particles change shape and speed up reactions right from your pocket.

What Did They Prove?

The researchers tested this assistant with two main "recipes" to see if it worked:

• The Shape-Shifting Test (Palladium): They asked the assistant to simulate Palladium particles under CO oxidation. The assistant correctly predicted that as the temperature went up, the particles would change from sharp, faceted shapes (like a diamond) to smooth, round shapes. This matched exactly what real scientists had seen in high-tech microscopes. It worked whether the data was already in the system or if ChatMOSP had to find it in a scientific paper first.

• The "Oscillation" Mystery (Platinum): They looked at Platinum particles, which are known to "dance" (oscillate) between high and low activity. The assistant simulated how the particle's shape changes based on gas pressure. It figured out the "secret loop":

  1. High gas pressure makes the particle round and active.
  2. Being active eats up the gas quickly.
  3. The gas pressure drops.
  4. The particle becomes sharp and slow again.
  5. Gas builds up, and the cycle repeats.

  ChatMOSP didn't just run the numbers; it explained why this cycle happens, connecting the shape changes to the speed of the reaction in a way that matches real-world experiments.

The Bottom Line

ChatMOSP isn't a magic box that invents new science on its own. Instead, it is a bridge. It takes the complex, specialized world of catalyst simulation and makes it accessible to anyone with a smartphone and a question. It ensures that the answers are still based on real physics (not just AI guessing), but it removes the barrier of needing to be a coding expert to get those answers. It turns a specialized scientific tool into a conversational partner for understanding how catalysts work in the real world.

Technical Summary

Technical Summary of ChatMOSP: A Chemistry-Grounded Mobile Agent for Working-State Catalyst Simulations

Problem Statement
Catalytic nanoparticles undergo dynamic restructuring under operando conditions, where their morphology and activity are governed by temperature, pressure, and gas composition. While physics-based simulations (specifically the Multi-scale Operando Simulation Package, or MOSP) can bridge the gap between reaction environments and nanoparticle behavior, their application remains limited to specialists. The core challenge is the "operational barrier": translating experimentally defined conditions (e.g., catalyst identity, temperature, pressure, gas composition) into the specific energetic, kinetic, and execution parameters required by MOSP. This translation demands specialized knowledge in computational catalysis, preventing routine use by experimentalists and non-specialists. Furthermore, existing scientific agents often automate predefined workflows or coordinate general tools but lack the capability to address specific, evolving catalytic science problems where structure and activity are coupled with the reaction environment.

Methodology
The authors introduce ChatMOSP, a chemistry-grounded mobile scientific agent designed to translate natural-language and voice-expressed catalytic requests into parameter-validated simulations. The system operates through the following technical framework:

• Interface and Architecture: Users interact via a mobile interface (Feishu) using text or voice. The system employs an OpenClaw-based hierarchical agent utilizing the GLM-5 language model for understanding and task coordination.
• Skill-Mediated Execution: A central design principle is that the language model is restricted to task interpretation and workflow coordination; it does not perform scientific predictions directly. Instead, it maps chemically meaningful variables onto Multi-scale Structure Reconstruction (MSR) and Kinetic Monte Carlo (KMC) task schemas required by MOSP.
• Parameter Handling:
  • Database Retrieval: When parameters are available, the agent retrieves them from a built-in MOSP database.
  • Literature-Assisted Construction: When parameters are missing, the agent initiates an automated literature-retrieval workflow. It constructs search queries (e.g., metal + reaction), filters results from open-access journals, extracts relevant energetic and interaction parameters from supplementary information, and formats them for MOSP execution.
• Validation Loop: The agent generates validated inputs, executes the MSR/KMC simulations, and returns results. The physical outputs (morphology, surface coverages, kinetic descriptors) are generated solely by the underlying MOSP physical models, ensuring physical consistency regardless of the language model backbone used.

Key Results
The paper validates ChatMOSP through three primary demonstrations using CO oxidation on Pd and Pt nanoparticles:

1. Model Preservation: ChatMOSP successfully reproduced temperature-dependent Pt morphology evolution trends previously established by expert-operated MOSP workflows, confirming that the agent acts as a model-preserving interface rather than an independent generator.
2. Database-Guided Reproduction (Pd): Using built-in parameters, ChatMOSP simulated Pd nanoclusters under CO oxidation. It accurately captured the experimentally observed temperature-induced transition from faceted morphologies (at lower temperatures, e.g., 473 K) to rounded morphologies (at higher temperatures, e.g., 873 K), consistent with in-situ TEM observations by Chee et al.
3. Literature-Assisted Parameterization (Pd): In a scenario where internal parameters were deliberately removed, ChatMOSP successfully retrieved Pd-CO/O parameters from online literature (specifically Nature Communications). Using these extracted values, it reproduced the same qualitative faceting-to-rounding evolution trend, demonstrating its ability to extend simulation capabilities beyond built-in databases.
4. End-to-End Mechanistic Insight (Pt): For Pt CO oxidation, the agent performed a closed-loop study linking local CO pressure to morphology and activity.
   • At low CO pressure (150 Pa), the system generated a faceted, low-activity morphology.
   • At high CO pressure (1500 Pa), it generated a rounded, high-activity morphology with a turnover frequency (TOF) over 10 times higher.
   • By analyzing these outputs, ChatMOSP reasoned a feedback mechanism explaining oscillatory CO conversion: high CO pressure stabilizes a rounded, high-activity state; rapid CO consumption lowers local pressure, causing a transition back to a faceted, low-activity state, which allows CO accumulation to restart the cycle.

Significance and Claims
The paper positions ChatMOSP not merely as a mobile interface for simulation software, but as a physically constrained mobile agent that makes working-state catalyst simulations accessible and interpretable. Its significance lies in:

• Lowering the Barrier: It converts chemically expressed reaction conditions into executable, physically validated simulations without requiring users to possess specialized knowledge of multiscale morphology and kinetic models.
• Bridging Experiment and Theory: It enables the direct translation of experimental observables (temperature, pressure, gas composition) into atomistic and mesoscale structural features.
• Mechanistic Interpretation: It demonstrates the capability to perform end-to-end studies that link environmental conditions to morphology evolution and catalytic performance, providing a physically interpretable feedback picture for complex phenomena like oscillatory reactions.
• Robustness: It establishes that catalyst-state predictions are governed by the underlying physical models (MOSP) rather than the language model, ensuring scientific rigor while leveraging the accessibility of natural language interfaces.