MLFF_CGMD: coarse-grained polymer MD trajectories

Langevin-dynamics trajectories of single coarse-grained bead-spring polymers in implicit solvent, generated with LAMMPS (LJ reduced units throughout). Intended as training data for machine-learned force fields / ML models of coarse-grained polymer dynamics.

All code, the LAMMPS input script, and the exact sampling configuration used to generate this dataset are at github.com/ljding94/MLFF_CGMD.

Repository layout

Path	Campaign	Runs
multi_topo/{topology}_run{N}.tar.gz	Multi-topology campaign (main)	5 topologies x runs 0-499 = 2500
linear_L100_box40_run{N}.tar.gz (root)	Legacy linear-only campaign	runs 0-999, chain length fixed at 100

Multi-topology campaign

One tarball per run, named {topology}_run{N}.tar.gz with topology ∈ {linear, ring, brush, star, dendrimer} and N ∈ [0, 499].

Every run independently samples all of its parameters (structural, interaction strength epsilon, bending stiffness kappa) uniformly from the ranges below, using a deterministic seed (see Reproducibility). The values actually used are recorded in each run's parameters.json.

Shared by all topologies:

Parameter	Range	Meaning
epsilon	[0.0, 2.0]	LJ well depth for all nonbonded bead pairs (self-attraction)
kappa	[0.0, 10.0]	cosine angle stiffness, E = kappa (1 + cos theta)
box_size	40 (fixed)	cubic periodic box edge

Structural parameters (integers sampled inclusive):

Topology	Parameters	Description
linear	chain_length [50, 200]	open chain of N beads
ring	ring_length [50, 200]	closed loop (bond between first and last bead)
brush	backbone_length [30, 80], grafting_density [0.1, 0.5], side_chain_length [3, 10]	linear backbone with side chains grafted at ~`grafting_density` fraction of backbone beads, evenly spaced
star	arm_length [10, 50], num_arms [3, 8]	num_arms linear arms bonded to one central core bead
dendrimer	generations [2, 4], branching_factor [2, 3], spacer [2, 6]	tree built outward from a core bead; each branch segment is spacer beads; the core spawns branching_factor branches, later branch points spawn branching_factor - 1

Contents of each tarball

{topology}_run{N}/
├── parameters.json        # every parameter of this run (sampled + fixed + seed)
├── polymer_{topology}.data  # LAMMPS datafile of the initial configuration
├── log.lammps             # full LAMMPS log (thermo output every 100 steps)
├── coord/dump.*.txt       # 1001 trajectory frames (see format below)
└── final_state.data       # LAMMPS datafile of the final configuration

parameters.json example (star, run 0):

{
  "mode": "rand", "topology": "star",
  "arm_length": 15, "num_arms": 4,
  "box_size": 40.0,
  "epsilon": 1.546645536562463, "kappa": 6.086129988363434,
  "timestep": 0.01, "equilibration_steps": 10000,
  "production_steps": 100000, "dump_every": 100,
  "run": 0, "seed": 3067890
}

Trajectory frames are LAMMPS dump custom text files written every 100 steps of the 100,000-step production run (frames at step 0, 100, ..., 100000). Columns:

id type x y z xu yu zu

x y z are wrapped coordinates, xu yu zu unwrapped; all beads are type 1 — the molecular topology (bond table) is in the .data files.

Simulation protocol

Coarse-grained bead-spring model, LJ units, all beads mass 1:

• Bonds: FENE, K = 30, R0 = 1.5, sigma = 1, epsilon = 1 (special_bonds fene)
• Angles: cosine, E = kappa (1 + cos theta), on all bonded bead triplets (branch points get one angle per neighbor pair)
• Nonbonded: LJ, sigma = 1, epsilon sampled per run, cutoff 2.5, potential shifted to zero at the cutoff
• Dynamics: Langevin thermostat, T = 1.0, damping 1.0, with fix nve; timestep 0.01
• Workflow: random-walk initial configuration → 10,000 equilibration steps with nve/limit 0.1 (removes overlaps; not dumped) → timer reset → 100,000 production steps, dumped every 100 steps

Reproducibility

Every run is fully determined by its (topology, run index): seed = 67890 + topology_offset + run, with offsets linear 0, ring 1e6, brush 2e6, star 3e6, dendrimer 4e6. The seed drives both the parameter sampling and the initial-configuration generation (the seed is also stored in parameters.json). Regenerate any run with the code repository:

python run_simulation.py --mode rand --topology star --box_size 40 --run 0

Legacy linear campaign (repo root)

1000 runs, linear_L100_box40_run{N}.tar.gz, N in [0, 999]: fixed chain_length = 100, box 40; only epsilon [0, 2] and kappa [0, 10] were sampled (seed = 67890 + N). Same model, protocol, and per-tarball layout; parameters.json there predates the multi-topology schema (no topology or run-length fields).

Example: download and extract one run

hf download ljding94/MLFF_CGMD multi_topo/star_run0.tar.gz \
    --repo-type dataset --local-dir .
tar -xzf multi_topo/star_run0.tar.gz