Citation Map¶

This page is the authoritative crosswalk between OpenLithoHub design decisions and the published works that justify them. If you find a constant, parameter, or schema rule with no matching row here, that's a documentation bug — please open an issue.

The BibTeX entries are stored in docs/references.bib. Citation keys below match that file verbatim.

How to read this table¶

Decision — what OpenLithoHub does (a constant, a default, a schema rule, a metric formula).
Where it lives — the file:line that implements it. Search for the citation key in source comments to find the inline justification.
Citation key — the BibTeX key in references.bib.
Section / claim — the specific section, table, or figure of the paper that backs the choice.

Forward-simulation and printability metrics¶

Decision	Where it lives	Citation key	Section / claim
Resist threshold defaults to `0.225`	`src/openlithohub/simulators/base.py`	`Yang2023_LithoBench`	§3.2 — calibrated against the ICCAD-16 reference resist model.
Forward-sim gate at submit-time (`l2_error_pixels` is required)	`src/openlithohub/leaderboard/tracker.py`	`Yang2023_LithoBench`	Table III — academic OPC printability = L2 + PVB on the simulated wafer image; an EPE-only score with no L2 is rejectable.
Hopkins SOCS uses 24 kernels by default	`src/openlithohub/simulators/hopkins_sim.py`	`Yang2023_LithoBench`, `Cobb1995_FastSparse`	Yang §3.2 / Table II for the count; Cobb for the SOCS construction itself.

Datasets¶

Adapter	Where it lives	Citation key	Notes
`LithoBenchDataset`	`src/openlithohub/data/lithobench.py`	`Yang2023_LithoBench`	NeurIPS'23 — paper introducing the benchmark consumed by this adapter.
`Iccad16Dataset`	`src/openlithohub/data/iccad16.py`	`Yang2016_ICCAD16Bench`, `Banerjee2013_ICCAD`, `Yang2020_BatchAL`	The 7nm-N7M2EUV release (Yang2016) extends the original ICCAD-2013 contest format (Banerjee2013). The N7M2EUV stack and per-layer mapping convention are documented in `Yang2020_BatchAL` §III-A.
`GanOpcDataset`	`src/openlithohub/data/ganopc.py`	`Yang2018_GANOPC`	DAC'18 — paper releasing the underlying mask-optimization dataset.

Models¶

Component	Where it lives	Citation key	Notes
`NeuralILTModel` (U-Net + L2/PVB co-loss)	`src/openlithohub/models/neural_ilt.py`	`Jiang2020_NeuralILT`	ICCAD'20 — architecture and loss formulation. Architecture audit is task 3.3.

Baselines¶

Component	Where it lives	Citation key	Notes
`batch_active_select` (uncertainty + diversity batch sampler)	`src/openlithohub/baselines/hotspot_batchal.py`	`Yang2020_BatchAL`	TCAD'20 §III — Eq. (8) uncertainty + Eq. (9) inner-product diversity. Greedy max-min selection replaces the paper's QP relaxation (Theorem 1 bounds the gap). The full active-learning loop (paper §3.4) is not shipped — see Candidate techniques table below.

Metadata format¶

Surface	Where it lives	Citation key	Notes
`DatasetAdapter.to_croissant()`	`src/openlithohub/data/base.py`	`MLCommons2024_Croissant`	MLCommons Croissant 1.0 JSON-LD format — the de-facto ML metadata schema (HuggingFace, Kaggle, Google).

Tile / halo strategy¶

Decision	Where it lives	Citation key	Notes
Process-node-aware halo sizing (`halo_px = max(ceil(OIR_nm/pixel_nm), receptive_field_px)`)	RFC 0005 (`docs/rfcs/0005-process-node-halo-sizing.md`), `src/openlithohub/workflow/halo.py`	—	Single-resolution physical optical-interaction-radius formula; no published-paper citation drives the formula itself (`OIR ≈ 10 × λ/(2·NA)` is textbook Hopkins/SOCS).

Candidate techniques (cited but not yet implemented)¶

The entries below are kept in docs/references.bib because they are plausible techniques for a future v0.x performance pass, not because the current code uses them. Adding a "Where it lives" pointer for any of these requires implementing the technique first.

Citation key	Technique	Where it would land if implemented	Status
`Yu2014_AccelerationOPC`	Coarse-to-fine multi-resolution SOCS forward-sim	`src/openlithohub/simulators/hopkins_sim.py`	Not implemented as of 2026-05-23. RFC 0005's halo pipeline uses a single-resolution OIR formula, not Yu2014's coarse-then-refine strategy. Verified against the actual code.
`Yang2020_BatchAL`	Full hotspot active-learning loop (detector training + lithography-simulation oracle alongside the §III sampler)	`src/openlithohub/baselines/hotspot_batchal.py` already ships the §III sampler; the loop would land beside it as `hotspot_al_loop.py`.	Sampler shipped 2026-05-23; full loop not implemented because OpenLithoHub does not ship a hotspot detector and the on-disk ICCAD16 corpus has only one testcase. See `out/plans/external-resource-utilization.md` Task #1 v0.2. Note: the same citation is also wired in for two unrelated purposes — the N7M2EUV stack / layer-mapping convention used by `Iccad16Dataset` (Datasets table) and the sampler itself (Baselines table).

Adding a new citation¶

Add the @type{key, ...} block to docs/references.bib. Use the FirstAuthor<YEAR>_ShortTopic key style.
Reference the key verbatim in the source comment / docstring at the point of use (so grep finds both sides).
Add a row above pointing at the file/section.
If the paper supersedes an existing citation, update the rows that pointed at the old key — don't leave stale pointers.

For the BibTeX file format and snapshotting policy, see References.