Terminology

A glossary of the acronyms and terms used across these docs. Acronyms also carry their expansion as a tooltip throughout the site: hover over an acronym such as SLE or MCMC on any page to see what it stands for.

Acronyms

Acronym	Stands for	What it means here
FUSION	the project name	The ice-fusion library (PyPI name ice-fusion, import name fusion). It scores an ice-sheet model ensemble against observations and weights each member's projection by historical skill.
SLE	sea-level equivalent	The contribution an ice volume would make to global mean sea level. The projection is an SLE distribution at the target year.
PSU-ISM	Penn State University Ice Sheet Model	The ice-sheet model whose ensemble members FUSION scores.
HBM	Hierarchical Bayesian Model	The structure of the inference: per-stream parameters share priors and combine into member weights. The prototype is PSUISM_HBM_V1.
MCMC	Markov chain Monte Carlo	The sampling method behind the posterior. Member weights are estimated by MCMC, so they are only trustworthy once the chains converge.
NUTS	No-U-Turn Sampler	The specific MCMC algorithm PyMC runs.
ESS	effective sample size	How many effectively-independent posterior draws inform a quantity. Reported as min_ess_bulk and min_ess_tail in the diagnostics.
PyMC	(library name)	The Python probabilistic-programming library that builds and samples the model.
VAF	volume above flotation	The physically correct basis for an SLE reduction. The v1 projection is a placeholder and does not yet compute it.
RMSE	root-mean-square error	The elevation uncertainty field is absolute_elevation_rmse.
NetCDF	Network Common Data Form	The on-disk format for the gridded observation and ensemble inputs (*.nc).
HPC	high-performance computing	Cluster login nodes are a common reason to relocate the cache with FUSION_CACHE.
RNG	random number generator	Matching RNG-call order between FUSION and the baseline is what keeps subsampling reproducible.
NSIDC	National Snow and Ice Data Center	Source of the region basins (MEaSUREs NSIDC-0709 v2).
MEaSUREs	Making Earth System Data Records for Use in Research Environments	The NASA program providing the region basins.
IMBIE	Ice sheet Mass Balance Inter-comparison Exercise	The basin definition behind the imbie_basins regions option.
AWS	Amazon Web Services	Source.Coop stores the observation bundles on AWS S3.
S3	Amazon Simple Storage Service	The object store the observation bundles live in (s3://...).
CLI	command-line interface	The AWS CLI is one of the two upload methods for the obs bundle.
API	application programming interface	The public Python API is whatever fusion/__init__.py re-exports.
OIDC	OpenID Connect	The mechanism behind PyPI Trusted Publishing; no API token is stored.
PyPI	Python Package Index	Where ice-fusion is published.
YAML	YAML Ain't Markup Language	The config file format load_config reads.
ISO	International Organization for Standardization	Obs bundle versions use ISO dates (YYYY-MM-DD).

CI is overloaded

"CI" means two different things in these docs, so it is left out of the hover tooltips. In the context of results it is the credible interval (e.g. the 5–95% interval on the projection). In the context of development and releases it is continuous integration (the GitHub Actions checks). The surrounding text always makes clear which one is meant.

Other terms

Term	What it means
R-hat (Gelman–Rubin)	A convergence diagnostic comparing between-chain and within-chain variance. A value above 1.01 means the chains disagree. Reported as max_rhat.
Divergence	A transition the sampler could not integrate accurately. Any divergences (n_divergences > 0) can bias the posterior, not just add noise.
Posterior mean	The headline member weight: the mean of a member's Bayesian weight w over the posterior. Weights sum to 1 across members.
Point estimate	The prototype's plug-in weight: a softmax of per-member log-likelihood at the posterior-mean parameters. Used as a cross-check on the posterior mean.
Credible interval	A Bayesian interval on a quantity. The projection is reported as a median with a 5–95% credible interval.
Stream	An observation type fed to the metric: elevation (thickness rate) or velocity. The two streams combine via stream_weights.
Plug-in weights	Deterministic weights computed by plugging the posterior-mean parameters into the likelihood, rather than averaging over the posterior.
Bit-exactness	Reproducing the prototype's arrays exactly (np.array_equal), not just within a tolerance. The validation harness checks this at Layers 1 and 2.
Science territory	Modules that port the prototype's science verbatim (inference/model.py, data/prepare.py). Their pinned defaults must not change without re-validating.