autoMD: Autonomous Simulation Control via Multi-Agent Systems

An industrial-grade agentic framework built on the Model Context Protocol (MCP) that automates end-to-end molecular dynamics (MD) workflows in LAMMPS. The system replaces traditional, brittle trial-and-error simulation loops with a tool-using AI controller backed by a deterministic auditor agent.

---

Architecture Overview

autoMD uses a decoupled multi-server architecture powered by FastMCP that communicates over a shared file system.

1. Pre-Processing Server

• Purpose: Topology manipulation, structure preparation, and dynamic input generation.

• Key Capabilities: Parses LAMMPS data files, performs atom deletion/renumbering for custom pore geometries, and injects runtime parameters into simulation templates.

• Supported Material Libraries: Includes literature-grounded and first-principles Lennard-Jones parameters for:

• Graphene (carbon-water interactions)

• h-BN (DREIDING-A force field parameterization)

• $\text{MoS}_2$ (UFF parameters for metal sites + alternating hourglass chemistry)

• $\text{Ti}_2\text{C}$ MXene

2. LAMMPS Runner Server

Purpose: Asynchronous simulation execution, hardware management, and process isolation.

• Key Capabilities: Launches detached LAMMPS jobs. Auto-detects local hardware acceleration options (MPI parallelization on CPU or Kokkos GPU targets) to optimize node footprint.
• State Protection: Operates with an independent Auditor Agent that intercepts tool commands to enforce hard safety boundaries (e.g., preventing timestep overreach, unphysical pressure ceilings, and simulation NaN/runaway instabilities).

3. Post-Processing Server

Purpose: Domain-specific calculations to evaluate reverse-osmosis (RO) performance.

• Key Capabilities: Streams/tails active simulation log files to parse thermodynamic streams, calculate real-time ion rejection rates , and compute total water flux curves ($L \cdot cm^{-2} \cdot day^{-1} \cdot MPa^{-1}$).

---

Core Tool Pipeline

Server	Core Tool	Functional Description
Pre-Processing	delete_atoms_and_rewrite	Modifies active coordinates and re-maps consistent topologies.
	reconstruct_full_filter	Rebuilds symmetric filter blocks from piston architectures.
LAMMPS Runner	start_lammps_detached	Launches simulations asynchronously to decouple LLM token life from execution times.
	get_lammps_status	Tracks process IDs and scans outputs for duplication or failure loops.
Post-Processing	desalination_water_flux	Extracts water permeation metrics from underlying trajectories.
	desalination_ion_rejection	Measures ion leakage behavior on the downstream membrane side.

---

Production Deployment & Validation

• in progress

---

Future Roadmap: Constraint-First Optimization

The project is scaling toward an autonomous discovery system for advanced membrane architectures. I want to a orchestrate a Constraint-First Optimization layer:

• Hard-Boundary Evaluation: Treats ion rejection percentage as a non-negotiable threshold parameter, rather than a soft weight in a loss function.

• Two-Phase Search Strategy:

1. Space-Filling Exploration: Initial broad sampling of the multi-material design space to identify stable operational baseline zones.

2. Bandit/Bayesian Refinement: Downstream optimization loops that maximize water flux within the validated structural safety bounds discovered in Phase 1.