Engineering stable light sources for bands nature does not supply — starting from EUV (13.5 nm) for semiconductor lithography, generalizing to any hard-to-generate wavelength lacking a stable natural emitter. The scope is the source problem, not one wavelength: EUV is the first instance, not the boundary.
"Is there really no path beyond 13.5 nm EUV?" — the framing question (from a Korean EUV analysis, Prof. Kwon Seok-jun) said there is "no clear technical alternative." lumen's verified-hypothesis lab reached a different conclusion:
The next-generation source problem is not a wavelength impossibility — it is an engineering / volume problem, reopenable down to one thermodynamic floor.
Framing claim lumen verdict (49 verified hypotheses)
───────────────────────── ───────────────────────────────────────
"no alternative beyond EUV" → wavelength has NO floor (H_011: λ ~ 1/E²)
→ the wall is FLUX, and flux is NOT a physics
ceiling (H_016: reopenable via ERL+FEL)
→ the terminal wall is economic, reopenable down
to the M8 waste-heat floor (H_019)
- Wavelength has no physics floor (H_011: λ ~ 1/E²) — an accelerator + undulator tunes 13.5 / 6.5 / 5 / 3 nm by an electron-energy dial (H_004/H_006/H_007), no emitter-element swap (Sn→Gd/Lu/La) + multilayer swap (Mo/Si→La/B) needed. But the dial is bounded by beam slice energy spread (H_031/032): only a beam compressed to ≪0.1% slice lases at 13.5 nm — which a conventional ERL achieves (FLASH, 2006) but a raw ~0.5% LPA beam does not (demonstrated LPA-FEL floor ~25 nm). Breakthrough (H_033): that slice-spread wall is FEL-specific — inverse-Compton scattering sidesteps it (no gain condition), reaching 13.5 nm with only ~2.2 MeV (≈455× less beam energy), slice spread merely broadening the line; the compact path reopens, the wall relocating to flux (reopenable, H_016). Honest residual: ICS-EUV at litho power is undemonstrated.
- The wall is flux (average power), not wavelength (H_008/H_005), and flux has no brightness/Liouville ceiling for a coherent source (H_016) — only efficiency + waste-heat (the M8 floor) survive.
- The answer: an accelerator-driven FEL EUV source — and, after a full real-data reference-match (H_030/031/032), its funded and physically-coherent form is a SHARED CONVENTIONAL energy-recovery linac (ERL) FEL amortized across many scanners (the xLight/KEK shape), NOT the tabletop laser-plasma (LPA) module-array first emphasized. Why the relocation: (1) the funded driver is a conventional ERL — zero funded programs use an LPA driver (H_030); (2) the win is M9-amortization (one ~$260 M / 10 kW source → ~$26 M/scanner at fan-out ~10, ~8× < LPP; it inverts at N=1), not the M7 module-learning curve, whose cryomodule Wright slope is shallow ~0.95 (H_031); (3) the wavelength "dial" is bounded by slice energy spread — a real ~0.5% LPA beam sits ~5× over the FEL gain criterion, single-pass cooling is undemonstrated, dechirping is not cooling, and the last orthogonal escape (a transverse-gradient undulator) is undemonstrated and breaks the LPA emittance budget (H_032); demonstrated LPA-FEL floor ~25–27 nm, never 13.5 nm. The conventional ERL clears that slice wall by RF compression (FLASH lased 13 nm, 2006). The LPA combination (H_022) stays a coherent longer-horizon variant; the concept itself is not novel (Nakajima 2014).
- Economics is a conjunction (H_019): a compact coherent source beats LPP $/wafer-layer only with efficiency η ≥ ~4× and amortization across ≥ ~3 scanners — neither lever alone — and never below the M8 waste-heat floor.
- Novel breakthrough on the accelerator power wall — SSMB (H_043): Steady-State MicroBunching is a third accelerator architecture beyond compact-LPA and single-pass FEL — a storage ring micro-bunched at the radiation wavelength every turn, radiating coherently (P ~ N², ~1e6×) and, because the ring reuses the beam continuously, delivering that power CW (~1 kW at 13-14 nm, ≥10× the HVM floor). The mechanism is proof-of-principle demonstrated (Tsinghua/MLS, Nature 2021). It is the missing middle — coherent average power without the FEL's dumped beam — and breaks the accelerator power wall. Honest residual: still a ring (footprint wall remains, so it rivals the FEL on footprint, not the compact LPA), and 13.5 nm kW SSMB is undemonstrated.
- Generalization — EUV is the first instance, not the boundary (H_039, meta-law M10):
the dial λ ∝ 1/γ² makes one accelerator + undulator/ICS span THz (3.4 MeV) → EUV (511 MeV) →
water window (1.08 GeV) → hard X-ray (5.94 GeV) — 6.5 orders of magnitude by electron energy
alone. The bands lacking a stable natural emitter (THz gap, EUV, water window, hard X-ray, nuclear
γ) are exactly the engine's no-incumbent market, and the campaign's whole structure (dial free ·
flux the wall · conventional accelerator the robust form · compact-coherent terminal-thin) recurs
at every rung (
state/band-generalization.md). - Final source-architecture ranking (EUV-litho high-flux requirement; ARCHITECTURE.json tree): ① shared conventional ERL FEL (funded near-term answer, xLight/KEK · H_030) · ② SSMB (novel, demonstrated mechanism, coherent CW · H_043) · ③ compact 7-MeV SSMB-Compton (tabletop dream resurrected, ~1 m ring, slice-immune · H_044-046) · ④ LPP (incumbent) · ⑤ LDP · ⑥ compact LPA-FEL (terminal-thin) · ⑦ recombination laser · ⑧ non-accelerator (fails HVM). The two novel paths ②③ converge on one shared, physics-allowed, undemonstrated gate — steady-state micro-bunching at 13.5 nm — so a single experiment de-risks both (H_048).
- Where it ends (honest terminus): the closed-form lab is complete (H_049) — every wall is classified, all 50 hypotheses run deterministically, and the question reduces to one physical experiment (13.5 nm micro-bunching + resist materials), out of in-silico scope. Fresh 2025-26 arXiv work corroborates the conclusions, and the cooling lever the lab derived closed-form (H_024) is exactly what the latest SSMB frontier adds (OSC, arXiv:2512.09399 · H_050).
Modeled on anima's UNIVERSE/: pre-register → falsify → run → verbatim verdict,
with all shared physics in a single deterministic harness.
[ HYPOTHESES/cards/H_*.md ] ─pre-registered + ≥5 falsifiers─▶ [ tool/lumen_optics.py ]
│ │ deterministic, stdlib-only
└── registry [ HYPOTHESES/REGISTRY.jsonl ] ◀─verbatim verdict─┴─▶ [ state/hX/run.py → result.json ]
HYPOTHESES/REGISTRY.jsonl— one JSON line per hypothesis (id · tier · verdict · artifacts).HYPOTHESES/cards/H_*.md— verifiedH_0xx(🟢, run to verdict) and abstractH_A*/H_B*(🜂, unverified conjecture with a falsifiable prediction, tier-separated).tool/lumen_optics.py— shared physics harness (mirror throughput, NA↔DoF, LPP power, undulator/inverse-Compton, ERL, cost-of-ownership, falsifier ledger).state/— per-hypothesis runs + result.json, pillar notes, brainstorms, meta-laws.
Status (computed by tool/qa.py, never hand-copied): 49 verified 🟢 · 1 falsified 🔴 · 10
abstract 🜂, all deterministic (each re-run is byte-identical and verdict-matched to the registry).
Key inputs are sourced against published data, as-of 2026-06 (state/sourced-parameters.md,
H_026) — core values real (optics R, LPP power, ~0.02% wall-plug, EUV-tool CAPEX), three optimistic
links honestly corrected (all "harder, not easier"). The abstraction lane peeled meta-laws M1–M10
(state/euv-meta-laws.md). Run python3 tool/qa.py to re-verify counts + determinism at any time.
python3 state/<hX>_<slug>_<date>/run_*.py # run any hypothesis (deterministic, $0 local)
python3 serve.py # view ARCHITECTURE.json in a browserlumen/
├─ HYPOTHESES/ — verification lab (REGISTRY.jsonl + cards/ + CLAUDE.md)
├─ tool/ — shared runnable physics harness (lumen_optics.py)
├─ state/ — per-hypothesis runs · pillar notes · meta-laws · brainstorms
├─ ARCHITECTURE.json — design SSOT (JSON tree · architecture.html viewer)
└─ CHANGELOG.md — history (append-only)
Verified H_0xx are closed-form public-physics relations with representative
parameters (each card lists its honest limits); they bound requirements and
scaling, not vendor performance. Abstract H_A*/H_B* and meta-laws are unverified
coordinates (🜂), not claims.
Funded-reality reference-match (H_030, the strongest honesty check): the accelerator-driven FEL EUV direction is real and funded — xLight (≤$150M CHIPS LOI, 2028 prototype, one source → ~20 scanners) and KEK (10 kW ERL EUV-FEL) — validating lumen's direction. But three honest corrections land: the funded driver is a conventional ERL, not the tabletop LPA this campaign emphasized; the funded economic lever is M9-amortization (shared source across scanners), not the M7 module-array; and 13.5 nm-on-LPA + single-pass cooling are undemonstrated (LPA-FEL floor ~25 nm). The concept predates lumen (Nakajima 2014) — the synthesis is not novel. The robust restatement of the answer is xLight-shaped (a shared conventional-ERL FEL amortized across many scanners), with the tabletop-LPA + module-array + cooling form as the longer-horizon, less-proven variant.
Where the terminal wall actually sits (settled, H_020): the one true physical ceiling is M8 (the 2nd-law waste-heat floor) — but with real-units cost-of-ownership it is only ~0.5% of LPP cost-of-ownership, economically negligible. The binding wall is CAPEX-per-wafer — attacked by M7 (learning curve / mass-produced modules) and M9 (amortization across scanners). Not wavelength, not flux, not waste-heat: the next-generation source problem is, at bottom, a CAPEX / volume question.