AI workflow assistant

Multi-step ACE3P campaigns, such as a mesh convergence study, a gradient sweep for multipacting, or an Omega3P-to-Track3P pipeline, involve many Slurm jobs with dependencies between them. The S3DF workflow project automates these with the Maestro workflow engine and uses an AI assistant to generate the workflow specification from a plain-English description.

Source: https://github.com/lge0303/s3df-workflow

How it fits together

flowchart LR
    U["User: 'Sweep gradient 10 to 50 MV/m<br/>in Track3P for pillbox cavity'"] --> A["AI assistant<br/>knows ACE3P inputs,<br/>S3DF Slurm settings"]
    A --> Y["Maestro YAML<br/>steps + dependencies + sweep"]
    Y --> M["maestro run"]
    M --> S["Slurm jobs:<br/>meshconvert, omega3p,<br/>track3p x N, analyze"]
    S --> R["Plots and summary"]

The assistant does not replace the workflow engine or the scheduler. It writes the specification; Maestro resolves dependencies and submits jobs; Slurm runs them. The user reviews the generated YAML before anything is submitted.

Validated pipelines

Workflow	What it does	Result
Omega3P single solve	Pillbox cavity, 16 MPI ranks	Mode 1 at 1.3138 GHz, matches reference to 8 digits
FE order convergence	Three parallel Omega3P runs, orders 1 to 3, with a fan-in comparison step	Order 1 has 0.08% error; order 2 converged
Omega3P to Track3P	Eigenmode solve, then multipacting scan over field levels, then plot generation	Resonant particles and enhancement counter at 23 to 25 MV/m

The whole Omega3P-to-Track3P pipeline runs in about five minutes on the S3DF milano partition.

Example specification

A generated Maestro step for the solver looks like this:

- name: omega3p-solve
  description: Run Omega3P eigenmode solver
  run:
    cmd: |
      source /sdf/group/rfar/ace3p/ace3p.sh
      cd $(WORKSPACE)
      srun --overlap --cpu-bind=none -n 16 omega3p pillbox.omega3p
    depends: [meshconvert]
    nodes: 1
    procs: 16
    walltime: "00:30:00"

Parameter sweeps use Maestro's global.parameters block, which expands one step into one job per value. Fan-in steps that aggregate results depend on step_*.

Relationship to this documentation

The assistant's knowledge of ACE3P input files comes from the same information presented in this manual. As the keyword reference is generated from source, it becomes a reliable foundation for validating generated inputs. Keeping the documentation and the workflow tooling in step is a goal of both projects.