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.
Define the forward model with JAX (hub:
jax-computing→ sub:jax-core-programming).Build the probabilistic model in NumPyro (hub:
bayesian-inference→ sub:numpyro-core-mastery).Run NUTS sampling and diagnose convergence (hub:
bayesian-inference→ sub:mcmc-diagnostics).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.
Set up force fields and initial configurations (hub:
simulation-and-hpc→ sub:md-simulation-setup).Run production simulations (hub:
simulation-and-hpc→ sub:advanced-simulations).Analyze trajectories: RDF, MSD, viscosity (hub:
simulation-and-hpc→ sub:trajectory-analysis).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.
Extract architecture and equations (
research-suiteskill:research-paper-implementation).Implement in JAX or Julia (science-suite hubs:
jax-computing,julia-language).Validate against reported benchmarks (
research-suiteskill:research-quality-assessment).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.
Trigger
scientific-reviewskill inresearch-suitewith the paper (PDF/DOCX/text) and optionally the target journal name.The skill performs six-competency analysis (domain, methodology, critical thinking, communication, integrity, efficiency) and produces a
.docxreferee report with Confidential Comments to Editor.For internal scoring without the
.docxdeliverable, useresearch-quality-assessmentinstead.
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:
Stage 1 —
spark-articulator: rough idea → 3-to-5-sentence articulation.Stage 2 —
landscape-scanner: three-layer literature scan + Reviewer 2 pass.Stage 3 —
falsifiable-claim: claim + Heilmeier catechism + kill criterion.Stages 4-5 —
theory-scaffold: stepwise derivation → LaTeX formalism (delegates tononlinear-dynamics-expertorstatistical-physicistin science-suite when applicable).Stage 6 —
numerical-prototype: JAX solver + three validation passes (delegates tojax-pro/julia-pro/simulation-expertin science-suite).Stage 7 —
experiment-designer: DoE + instrument capability map (3× margin rule).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.