Scientific Workflows

Patterns for using the science-suite agents and hub skills in research computing pipelines.

Note

Since v3.1.0, skills use a two-tier Hub Skill architecture. The hub skills listed below route to specialized sub-skills via their Routing Decision Tree. You invoke the hub; it dispatches to the right sub-skill automatically.

Bayesian Inference Pipeline

A typical Bayesian parameter estimation workflow combines @jax-pro for model implementation with @research-expert (research-suite) for methodology.

  1. Define the forward model with JAX (hub: jax-computing → sub: jax-core-programming).

  2. Build the probabilistic model in NumPyro (hub: bayesian-inference → sub: numpyro-core-mastery).

  3. Run NUTS sampling and diagnose convergence (hub: bayesian-inference → sub: mcmc-diagnostics).

  4. Visualize posteriors with ArviZ (hub: ml-and-data-science → sub: scientific-visualization).

import jax
import numpyro
from numpyro.infer import MCMC, NUTS

# 1. Forward model (JIT-compiled)
@jax.jit
def model_predict(params, x):
    return params["a"] * jax.numpy.exp(-params["k"] * x)

# 2. NumPyro model
def bayesian_model(x, y_obs=None):
    a = numpyro.sample("a", numpyro.distributions.LogNormal(0, 1))
    k = numpyro.sample("k", numpyro.distributions.HalfNormal(1))
    sigma = numpyro.sample("sigma", numpyro.distributions.HalfNormal(0.1))
    y_pred = model_predict({"a": a, "k": k}, x)
    numpyro.sample("obs", numpyro.distributions.Normal(y_pred, sigma), obs=y_obs)

# 3. Run MCMC
kernel = NUTS(bayesian_model)
mcmc = MCMC(kernel, num_warmup=500, num_samples=2000, num_chains=4)
mcmc.run(jax.random.PRNGKey(42), x_data, y_obs=y_data)

Agent team: Use Team 13 (bayesian-pipeline) for multi-agent Bayesian workflows.

Molecular Dynamics Campaign

For MD simulation campaigns, combine @simulation-expert with @jax-pro for differentiable physics.

  1. Set up force fields and initial configurations (hub: simulation-and-hpc → sub: md-simulation-setup).

  2. Run production simulations (hub: simulation-and-hpc → sub: advanced-simulations).

  3. Analyze trajectories: RDF, MSD, viscosity (hub: simulation-and-hpc → sub: trajectory-analysis).

  4. Compute correlation functions (hub: correlation-analysis → sub: correlation-computational-methods).

Agent team: Use Team 14 (md-campaign) for coordinated MD workflows.

Research Paper Implementation

Reproducing results from published papers requires systematic methodology. In v3.4.0 the research-methodology skills moved from science-suite to the new research-suite.

  1. Extract architecture and equations (research-suite skill: research-paper-implementation).

  2. Implement in JAX or Julia (science-suite hubs: jax-computing, julia-language).

  3. Validate against reported benchmarks (research-suite skill: research-quality-assessment).

  4. Create publication-quality figures (science-suite hub: ml-and-data-science → sub: scientific-visualization).

Agent team: Use Team 16 (paper-implement) for coordinated reproduction workflows.

Peer Review of a Manuscript

Producing a rigorous, journal-ready peer review is a distinct workflow from reproducing papers or assessing quality internally.

  1. Trigger scientific-review skill in research-suite with the paper (PDF/DOCX/text) and optionally the target journal name.

  2. The skill performs six-competency analysis (domain, methodology, critical thinking, communication, integrity, efficiency) and produces a .docx referee report with Confidential Comments to Editor.

  3. For internal scoring without the .docx deliverable, use research-quality-assessment instead.

Research-Spark: Idea to Fundable Plan

Refining a rough research idea into a scoped, testable, fundable program. Eight-stage artifact-gated pipeline in research-suite:

  1. Stage 1 — spark-articulator: rough idea → 3-to-5-sentence articulation.

  2. Stage 2 — landscape-scanner: three-layer literature scan + Reviewer 2 pass.

  3. Stage 3 — falsifiable-claim: claim + Heilmeier catechism + kill criterion.

  4. Stages 4-5 — theory-scaffold: stepwise derivation → LaTeX formalism (delegates to nonlinear-dynamics-expert or statistical-physicist in science-suite when applicable).

  5. Stage 6 — numerical-prototype: JAX solver + three validation passes (delegates to jax-pro / julia-pro / simulation-expert in science-suite).

  6. Stage 7 — experiment-designer: DoE + instrument capability map (3× margin rule).

  7. Stage 8 — premortem-critique: failure narratives + simulated reviewers.

The research-spark-orchestrator agent drives the pipeline, owns _state.yaml, and fans out to parallel sub-agents at natural stage boundaries.