English · 简体中文

Molecular Dynamics, One API Away.

DPAstra = DPA + astra (“stars”).
DPA-native. API-first. Built for frontier MLIPs.

Run Molecular Dynamics Through an LLM-Style API

The destination: interact from a local Python notebook, run molecular dynamics with 10M+ atoms in the cloud, and stream live thermodynamics and trajectory previews—all through one API.

LLM APIs let researchers call frontier models without operating the inference stack. DPAstra applies the same abstraction to stateful molecular dynamics: scientists declare a potential, atoms, and a simulation protocol; qualified workers own the accelerator runtime and the complete MD session.

DPAstra is an independent API-first layer built to serve DPA models and other frontier MLIPs, designed around DeePMD-kit execution interfaces and the open DeepModeling ecosystem.

Install

Research Preview. The default PyPI install contains the client SDK only. For source development and local execution, see the development guide.

pip install dpastra

Models

The API model ID is a human-readable catalog alias, not an authentication credential. Users submit that ID; the service resolves it to one exact immutable model revision before the run is persisted. Advanced callers can optionally assert model_revision, while the ordinary model list and submission path stay digest-free. Catalog IDs may change before the first public release. — means no DPAstra-qualified entry.

[!WARNING] Research Preview. The API and qualified model catalog may change before the first public release. A deployment provider supplies the HTTPS base URL, API key, and models available to each user. Model availability is not proof of scientific accuracy or production convergence for a particular system.

_{Scale note: 10M+ atom cloud MD depends on future qualified LAMMPS/Kokkos and multi-GPU fleet backends; it is not a current validated capability.}

Use

[!WARNING] The API is under active development. This example shows the current intended SDK shape and may change before release.

Ask your deployment provider for DPASTRA_BASE_URL and DPASTRA_API_KEY, then keep both values in your environment rather than in a notebook or script. The SDK has no default service URL. Install the optional terminal dashboard with pip install 'dpastra[watch]' before calling run.watch().

import os

from ase.build import bulk
from dpastra import Client, MD

atoms = bulk("Cu", cubic=True)

try:
    client = Client(
        os.environ["DPASTRA_BASE_URL"],
        api_key=os.environ["DPASTRA_API_KEY"],
    )
except KeyError as error:
    raise SystemExit(f"Set {error.args[0]} before connecting to DPAstra") from None

with client:
    run = client.md.submit(
        atoms,
        model="omat24-dpa4c-nano",
        dynamics=MD.verlet(timestep_fs=0.1),
        velocity_initialization=MD.zero_velocities(),
        steps=4,
        checkpoint_every=2,
        store_every=1,
        thermo_every=1,
        stream_every=1,
    )

    run.watch()
    run.download_artifact("trajectory", "trajectory.traj")

The deployment operator publishes the endpoint, credential, and available catalog. Users can discover model IDs with client.models() and select one for a run. DPAstra freezes the resolved exact revision in the durable run record; streamed frames are previews, while the downloaded trajectory is the authoritative result.

Documentation

• MD parameters — constant-temperature setup, timestep, velocities, constraints, cadence, and planned LAMMPS mappings.
• Watch a run — terminal and notebook previews, with the authoritative trajectory downloaded after completion.
• Development guide — current source setup and local execution.
• Scientific contract — units, ensembles, cadence, and supported boundaries.
• Architecture — control plane, workers, sessions, and artifacts.
• Product plan — evidence-gated development direction.

DPAstra is an independent Research Preview, not an official DeepModeling project. It is licensed under the Apache License 2.0.