PolyJarvis: LLM Agent for Autonomous Polymer MD Simulations

PolyJarvis is an LLM-driven autonomous agent that integrates with the RadonPy platform to execute end-to-end all-atom molecular dynamics simulations for polymers, successfully predicting key properties like density and bulk modulus with high accuracy while demonstrating that LLM agents can match expert-level performance in polymer simulation workflows.

The Gist

Imagine you want to build a new type of plastic, like a super-strong water bottle or a flexible phone case. To do this, scientists usually need to run complex computer simulations called Molecular Dynamics (MD). Think of these simulations as a high-speed, microscopic movie where they watch billions of tiny atoms bounce, stick, and wiggle to see how the material behaves.

The problem? Making these movies is incredibly hard. It's like trying to direct a Hollywood blockbuster where you have to:

1. Write the script (choose the right physics rules).
2. Cast the actors (build the molecule).
3. Set the stage (build the computer model).
4. Run the cameras (run the simulation).
5. Edit the footage (analyze the results).

Usually, only a PhD expert with years of training can do all this. If they make a tiny mistake in the script, the whole movie is garbage.

Enter PolyJarvis.

The "AI Director"

The researchers at Carnegie Mellon University built PolyJarvis, an artificial intelligence agent that acts like a super-smart, autonomous movie director for these atomic simulations.

Instead of a human typing out complex code, you just talk to PolyJarvis in plain English. You might say: "Hey, I want to know how strong Polyethylene (PE) is, or what temperature Polystyrene (aPS) melts at."

PolyJarvis then takes over the entire production:

• It reads the script: It looks up the chemical structure of the plastic you mentioned.
• It hires the crew: It automatically chooses the best "force field" (the rulebook of physics) for that specific plastic.
• It builds the set: It constructs the digital atoms and packs them into a virtual box.
• It runs the cameras: It launches the simulation on powerful supercomputers (GPUs).
• It edits the movie: It watches the simulation, checks if the atoms are behaving correctly, and calculates the final properties (like density or melting point).

How It Thinks (The "Brain" vs. The "Hands")

PolyJarvis uses a "brain" (a Large Language Model, specifically Claude) and connects it to "hands" (specialized software tools called RadonPy and LAMMPS) via a universal translator called MCP (Model Context Protocol).

• The Brain (Claude): It understands your request, makes decisions, and spots errors. If the simulation crashes because of a weird glitch, PolyJarvis doesn't just stop; it says, "Oh, the atoms are too crowded. Let me adjust the pressure and try again." It learns from its mistakes in real-time.
• The Hands (RadonPy/LAMMPS): These are the tools that actually do the heavy lifting of building molecules and crunching the numbers.

The Test Drive

The team tested PolyJarvis on four common plastics:

1. Polyethylene (PE): Used in plastic bags.
2. Polystyrene (aPS): Used in Styrofoam cups.
3. PMMA: Used in Plexiglass.
4. PEG: Used in lotions and medicines.

They asked PolyJarvis to predict three things: Density (how heavy it is), Bulk Modulus (how hard it is to squish), and Glass Transition Temperature (the point where it goes from hard to gooey).

The Results: A Mixed Bag, But Promising

• The Wins: For Polystyrene and PMMA, PolyJarvis got the density and "squishiness" almost exactly right, matching what human experts have found in labs. It even predicted the melting point of PMMA very accurately.
• The Struggles: For some plastics, the AI predicted they would stay solid at higher temperatures than they actually do. However, the researchers realized this wasn't because the AI was "dumb." It's because the computer simulation cools down the plastic too fast (like putting a hot pan in a freezer), which tricks the atoms into freezing in a weird shape. This is a known limitation of the physics, not the AI.
• The Self-Correction: In one case, the AI noticed the simulation was getting stuck in a bad state. It realized, "Hey, my chain length is too short," and automatically fixed the code to make the chains longer, improving the results.

Why This Matters

Before PolyJarvis, if you wanted to simulate a new plastic, you needed a team of experts and weeks of trial and error. Now, you can just ask an AI, and it will run the experiment for you, fixing its own mistakes along the way.

The Bottom Line:
PolyJarvis isn't replacing scientists yet, but it's like giving every materials scientist a tireless, hyper-intelligent intern who never sleeps, never gets tired of debugging code, and can run thousands of experiments while you go get coffee. It turns a process that used to require a PhD in computer science and chemistry into a simple conversation.

In short: PolyJarvis is the first AI that can not only talk about science but actually do the science, building and testing virtual materials from start to finish on its own.

Technical Summary

1. Problem Statement

All-atom Molecular Dynamics (MD) simulations are a powerful tool for predicting polymer properties (e.g., glass transition temperature $T_g$, density, bulk modulus) from molecular structure. However, their practical application is hindered by:

• High Expertise Barrier: Executing simulations requires specialized knowledge in force field selection, system construction, multi-stage equilibration, and property extraction.
• Reproducibility Issues: Manual, decision-intensive workflows often lead to divergent results for the same system due to human variability.
• Limitations of Existing Automation: Current tools (e.g., RadonPy, Polymatic) execute predefined protocols but lack the ability to make informed, context-aware decisions based on specific polymer chemistry. Machine learning models offer speed but often sacrifice physical interpretability.

2. Methodology: The PolyJarvis System

The authors present PolyJarvis, an autonomous agent that integrates a Large Language Model (LLM) with scientific simulation software to execute end-to-end polymer MD workflows from natural language inputs.

System Architecture

PolyJarvis operates via a Client-Server architecture mediated by the Model Context Protocol (MCP):

1. LLM Backbone: Uses Anthropic's Claude Sonnet 4.5 to process prompts, maintain workflow state, and make simulation decisions.
2. Local MCP Server (RadonPy): Wraps the RadonPy library for molecular construction, force field assignment, charge calculation, and property analysis.
3. Remote MCP Server (LAMMPS): Manages GPU-accelerated LAMMPS simulations on Lambda Labs cloud infrastructure (4× Quadro RTX 6000 GPUs).

Autonomous Decision-Making Workflow

Unlike static scripts, PolyJarvis makes intelligent, polymer-specific decisions at every stage:

• Classification & Force Field Selection: The agent classifies the polymer against 21 backbone classes (PoLyInfo scheme) to select appropriate force fields (e.g., GAFF2 vs. GAFF2 mod) and charge methods (RESP vs. Gasteiger).
• System Construction: It determines chain lengths and system sizes (targeting 6,000–15,000 atoms) to ensure statistical sampling and convergence of conformational statistics.
• Equilibration Protocol: The agent designs staged protocols (energy minimization, NVT heating, NPT compression/decompression, cooling) and adapts parameters (annealing cycles, compression methods) based on system behavior.
• Error Recovery: The agent autonomously diagnoses and resolves simulation errors (e.g., PPPM out-of-range atoms, pair style mismatches, velocity re-initialization bugs) by modifying scripts and re-running jobs.
• Property Extraction: It calculates $T_g$ via bilinear fitting of density-temperature curves, density via time-averaging, and bulk modulus via volume fluctuations.

3. Key Contributions

• First Fully Autonomous Polymer MD Agent: Demonstrates an LLM agent capable of executing validated all-atom MD simulations with quantitative experimental benchmarking, a capability previously unproven in this domain.
• Intelligent Workflow Orchestration: Moves beyond "hard-coded" heuristics to domain-knowledge-driven decision-making (e.g., selecting specific force field modifications for aromatic rings or adjusting equilibration stages based on initial run failures).
• Iterative Self-Correction: The system demonstrates the ability to refine protocols across replicates (e.g., increasing chain length for Polyethylene after Run 1, switching compression methods for PMMA after a crash).
• Integration of MCP: Successfully utilizes the Model Context Protocol to bridge LLM reasoning with complex, asynchronous scientific computing tools.

4. Results and Validation

The system was validated on four benchmark polymers: Polyethylene (PE), Atactic Polystyrene (aPS), Poly(methyl methacrylate) (PMMA), and Poly(ethylene glycol) (PEG).

Performance Metrics

Acceptance criteria were set as: $|T_{sim} - T_{exp}| \le 20$ K, density error $\le 5\%$, and bulk modulus error $\le 30\%$.

Polymer	Property	Result vs. Experiment	Status
PMMA	$T_g$	395 K (Exp: 377–385 K); +10–18 K error	Pass (Only PMMA passed strict $T_g$ criteria)
aPS	$T_g$	416 K (Exp: 353–373 K); +43–63 K error	Fail (Consistent with known MD cooling-rate bias)
PEG	$T_g$	260 K (Exp: 206–213 K); +47 K error	Fail (Consistent with MD cooling-rate bias)
PE	$T_g$	281 K	N/A (Experimental $T_g$ for amorphous PE is ambiguous)
aPS	Density	1.054 g/cm³ (Exp: 1.053 g/cm³); +0.1% error	Pass
PMMA	Density	1.124 g/cm³ (Exp: 1.18 g/cm³); -4.8% error	Pass
PEG	Density	1.099 g/cm³	Pass (Matches AA MD references)
PE	Density	1.069 g/cm³ (Exp: 0.855 g/cm³); +25% error	Fail (Due to kinetic trapping/protocol error)
Bulk Moduli	All Systems	Within $\pm 30\%$ of references	Pass (All 4 systems)

Summary of Validation:

• 5 out of 8 property-polymer combinations with direct experimental references met strict acceptance criteria.
• $T_g$ Limitations: The overestimation of $T_g$ in aPS, PMMA, and PEG is attributed primarily to the intrinsic cooling-rate bias of all-atom MD (simulations cool too fast compared to experiments), not agent error.
• Density Failure: The PE density overestimation (+25%) was identified as a specific agent failure where the protocol drove the system into a kinetically trapped, over-dense state.

5. Significance and Future Outlook

• Paradigm Shift: PolyJarvis demonstrates that LLM agents can shift the researcher's role from manual workflow execution to high-level ideation and experimental design.
• Robustness: The agent successfully resolved 13 distinct simulation errors across four categories without human intervention, showcasing resilience in complex computational environments.
• Future Work: The authors plan to expand force field coverage (OPLS-AA, TraPPE-UA), formalize protocol adaptation into a rigorous decision framework, increase ensemble sizes, and extend validation to semicrystalline polymers, copolymers, and blends.

Conclusion: PolyJarvis represents a significant step toward fully autonomous materials discovery, proving that LLM-driven agents can execute complex, physics-based simulations with results comparable to expert-run studies, provided the limitations of the underlying physics (like cooling rates) are understood.