Simulation Management Beyond “Run and Hope”: Adaptive Placement, Archiving, and SuperNu as a Realization

Status: Draft v3 — added motivation section per Richard’s guidance Date: 2026-05-10 Authors: Junior (drafting), based on research by coding-rift and code-op-astro agents

Motivation: Why Simulation Management Is a First-Class Problem

The traditional approach to simulation-driven science treats each simulation as an independent event: write the input, submit the job, wait for the output, move on. This works when simulations are few and the parameter space is small. It breaks down as workflows scale.

The placement problem. Gravitational wave parameter estimation involves exploring high-dimensional parameter spaces with expensive simulations. Numerical relativity (NR) waveforms — the gold standard for accuracy — take weeks to generate on hundreds of cores. Simply throwing compute at the problem is not sustainable. The question isn’t just “can we run more simulations?” but “which simulations should we run, and where?” Adaptive placement answers that by using prior results and workflow structure to guide decisions — routing new simulations to the nodes or backends best suited to the current state of the analysis.

The archiving problem. Even when simulations are run well, the outputs are often lost to “ad-hoc directory structures and institutional memory.” Six months later, recreating what was done, why, and what the results meant is archaeology, not science. The cost of re-running simulations — in compute, time, and reproducibility — makes proper archiving a prerequisite for long-term scientific productivity, not an afterthought.

The connection to gravitational wave science. These problems are amplified in GW parameter estimation, where:

The parameter space is high-dimensional and structured (masses, spins, eccentricity, sky location, distance).
Each simulation is expensive (NR waveforms, kilonova light curves, GRMHD merger simulations).
Results must be reused across analyses (waveform catalogs, surrogate models, population inference).
New LIGO/Virgo/Kamioka triggers demand rapid response — the system must route work efficiently without manual intervention.

Our group has pioneered adaptive simulation placement across several projects: targeted NR waveform banks for eccentric GW events (GW200208_22, GW200105), NRHybSur3dq8 surrogates for fast LISA/RIFT parameter estimation, and synthetic population validation to expose biases from circular-only PE. The same philosophy — placing simulations deliberately in parameter space to maximize information gain — extends naturally to kilonova radiative transfer (via SuperNu) and, in time, to GRMHD simulations of mergers with matter.

The Problem: Simulations Are Cheap to Start, Hard to Manage

Running a simulation is easy. Running the right simulation, in the right place, at the right time, and then finding it again six months later — that’s hard. As simulation workflows grow in complexity, the gap between “submitting a job” and “getting useful science out of the output” widens. Two concepts help close that gap: adaptive simulation placement and simulation archiving. The RIFT Simulation Manager API provides a framework for both; SuperNu provides a concrete implementation.

Adaptive Simulation Placement

What it means: Adaptive placement is about choosing where a simulation runs and how it’s configured based on available resources, prior results, and the structure of the parameter space. Instead of manually matching simulations to machines or tuning parameters by hand, the system reasons about the workflow as a whole.

Where it lives in RIFT: The RIFT Simulation Manager provides a structured backend API (gw_pe_synthetic) that decouples simulation logic from execution details. Backends can implement custom placement strategies — deciding which nodes to use, how to partition the workload, and when to adaptively refine parameter estimates.

Why it matters: Parameter estimation for gravitational wave events involves exploring high-dimensional spaces with expensive simulations. Adaptive placement keeps the pipeline responsive: new LIGO/Virgo triggers can be routed efficiently, and intermediate results can guide where to invest the next round of compute.

Simulation Archiving

What it means: Archiving isn’t just “saving the output” — it’s about preserving provenance, provenance chains, and searchability. When a simulation is archived properly, future users (or future versions of yourself) can understand exactly what was run, why, and what the results mean — without hunting through ad-hoc directory structures.

Where it lives in RIFT: The Simulation Manager’s documentation system encodes this philosophy. DESIGN.md serves as the single source of truth; automated Sphinx tooling converts it to documentation on every build. The backends/ tree records the full API surface, making it auditable and discoverable.

The Simulation Manager API: A Common Interface

The real power of the RIFT Simulation Manager isn’t just the backends themselves — it’s the API contract. Any code that implements the simulation manager interface can slot into the same workflow: RIFT’s parameter estimation, downstream analysis, archiving, and reuse. This means:

New simulation backends can be dropped in without rewriting the workflow.
Results from different codes can be compared in a common framework.
The system can reason about where a given simulation fits in the overall pipeline.

SuperNu as a Realization of the API

What is SuperNu? SuperNu is a Monte Carlo radiative transfer code (Fortran, LANL) that simulates time-dependent radiation transport in matter — the physics that determines what a supernova or kilonova looks like to an observer.

How it fits the pattern: SuperNu isn’t just a physics code someone wrote — it’s a concrete realization of the simulation manager concept:

It implements a simulation interface — SuperNu can be treated as a backend that takes structured input (density, velocity, composition from a hydro simulation) and produces structured output (light curves, spectra).
It enables adaptive post-processing — Different radiation transport strategies can be explored without re-running the underlying hydro.
It feeds the archival pipeline — Its output is exactly the kind of thing that needs to be preserved with full provenance: this input model produced this observable signature.

Technical details:

Implements Implicit Monte Carlo (IMC) and Discrete Diffusion Monte Carlo (DDMC) methods
Multi-frequency opacities: Thomson scattering, bound-bound, bound-free, free-free
1D/2D/3D geometries, static or homologously expanding grids
Hybrid MPI + OpenMP parallelization
Output: Light curves and spectra — the observables that connect simulation to observation

Why users care: When a gravitational wave event triggers follow-up observations, the chain from “detection” to “physical interpretation” runs through radiative transfer. SuperNu gives that chain a concrete form: take the output of a hydro simulation, feed it through the radiation transport layer, and get predictions for what JWST or other instruments should see. Making that pipeline explicit — via a simulation manager API — is what turns “we ran some code” into reproducible science.

How They Relate

Aspect	RIFT Simulation Manager API	SuperNu
Role	Defines the interface + implements adaptive placement & archiving	Realizes the interface for radiative transfer post-processing
Scope	Framework-level: parameter estimation workflows	Domain-level: supernova/kilonova light curves and spectra
Output	Parameter estimates, waveform catalogs	Observables: light curves, spectra
Connection	SuperNu plugs into the RIFT workflow as a backend	Takes hydro output → produces observable signatures

Files Referenced

RIFT Simulation Manager

Code: MonteCarloMarginalizeCode/Code/RIFT/simulation_manager/
Docs: ~/RIFT/research-projects-RIT/docs/source/api_reference/simulation_manager/
API: gw_pe_synthetic backend implementation
arXiv: RIFT paper

SuperNu

GitHub: rjw57/SuperNu — Monte Carlo radiative transfer
Origin: Los Alamos National Laboratory (LANL)

This is a draft v2 with the adaptive placement / archiving framing. Feedback welcome.

Simulation Management Beyond Run and Hope: Adaptive Placement, Archiving, and SuperNu as a Realization