Neil Mitchell has spent forty years watching steel move. His new arXiv retrospective, Large Magnet Structures for Fusion Technology: Lessons from ITER, is not a fusion-is-coming piece — it is a metallurgy confession. The load-bearing sentence is his own: new materials are “easy on a laboratory scale, but often not transferable to an industrial large scale.” Everything downstream of that — the 50-tonne 316LN forgings, the welds up to 0.3 m deep held under controlled distortion, the two-year schedule slip — is what happens when a lab result meets a real supply chain.

Start with the machine, per ITER Organization’s published specifications. Eighteen toroidal field coils, each 9 × 17 m and 330 t, roughly 190 t of high-strength 316LN stainless per case, plate up to 180 mm thick. Operating temperature ~4 K, cooled by supercritical helium; peak field 11.8 T; 41 GJ stored. Assembly tolerance to get a coil into the tokamak at all: a few millimetres, and under 1 mm where interfaces demand it.

Now the one number the whole piece turns on. Austenitic stainless contracts about 0.30% of its length between 293 K and 4 K — the NIST cryogenic fit for 304 gives an integrated contraction near −0.296%, about 3 mm per metre. Do the arithmetic Mitchell wants on the page: a 17 m coil cold-shrinks roughly 50 mm on the way to 4 K, about 27 mm across the 9 m width. The structure that must sit inside a few millimetres at operating temperature is therefore never that shape when it is built, machined, or measured. Every dimension on the shop drawing is a warm dimension for a cold machine, and the discipline is bookkeeping the difference.

←TODAY: A 17 m fusion coil moves 50 mm between the room it is built in and the field it works in, and still has to meet its neighbour inside a millimetre. →3012: The offices that survive specify the temperature a dimension was taken at, not just the dimension. Fulcrum: A tolerance is a promise about two states of the same part — you only see it once you hold both the warm and the cold in view.

The failures prove it. The vacuum-vessel sectors (2022–2023) accumulated weld distortion that only surfaced at the field joints between parts from two supply chains; a 485 t module was pulled from the pit for repair. The thermal shields (2021) cracked from stress-corrosion traced to chlorine residues trapped near welds during coating — found 18 months late, by helium testing. Neither was a physics failure — one a process residue, one accumulated distortion at an interface — and both were found late, on delivered work.

The Swiss clause: every ITER conductor sample — China, Korea, Europe, Japan, Russia, USA — was qualified at SULTAN, the PSI Villigen bench run by EPFL’s Swiss Plasma Center, the only facility that could do it. The bottleneck for the largest magnet ever built was a bench in Villigen.

Atelier: Your curtain wall is this problem with the temperature term turned down, not off — an Alpine facade swings tens of millimetres a year between a February install and an August load. Mitchell’s two named remedies are ones your site already half-uses: overmetal (leave material on, machine to fit once the neighbour is real) and shimming (design the adjustment joint in, on purpose). And SIA 414/1 and 414/2 say plainly that construction is not precision manufacturing — the espazium commentary frames tolerance as a Verständigungsgrundlage, the shared language in which planner and executor agree what “close enough” means before anyone measures. Monday move: on your next thick-section or long-span detail, write the reference temperature next to every critical dimension and name who owns the adjustment joint.

Hack: Compute the cold-short before you trust a warm drawing. Convert a length and a contraction fit into the millimetres a part actually loses — the move ITER pays a schedule slip to skip. Swap the coefficient for your material’s real integrated contraction over your temperature swing; the part only ever exists as two dimensions, and you must carry both.

L = 17.0                    # warm length of the TF coil, metres
contraction = 0.00296       # NIST 304-SS fit, 293 K to 4 K
shrink_mm = L * contraction * 1000
print(f"cold-short: {shrink_mm:.0f} mm")   # ~50 mm to land

Read the error bar before the headline. Mitchell’s paper is dull in the best way — it costs its lessons in forgings and pit removals so you can buy them for a coefficient. Before your next detail meets its neighbour, decide whether you are buying overmetal or shims, and at what temperature.

Source: arXiv