Industries / Frontier science / Fusion energy
Materials
This is the work of choosing and testing the solids that face the plasma, hold the machine together and carry the magnets' current, all of which must survive heat, fast particles and neutrons for years.
6
Milestones achieved, from 1997 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1997
Carbon walls soak up tritium
The two tokamaks that burned deuterium-tritium fuel in the 1990s, TFTR in Princeton and JET in England, both had carbon walls, and both kept much of their tritium in them: eroded carbon settled back onto surfaces and buried fuel with it, a process called co-deposition. TFTR's walls still held 51% of the 5.2 g of tritium it had injected after later deuterium-only running, and JET's held 17% of about 35 g after its 1997 campaign (DTE1) and clean-up pulses. In both machines, when the stored tritium neared the site's safety limit, staff had to stop and run long removal campaigns. A 2001 Princeton Plasma Physics Laboratory report comparing the two called such retention rates "unacceptable in a next step device and in a fusion reactor".
Lessona wall that traps fuel faster than it can be cleaned out cannot be used in a power plant, which pushed later machines toward metal walls.
Credit
the TFTR and JET teams; comparison by C. H. Skinner and G. Federici
2008
Helium turns tungsten fuzzy
Tungsten is the wall armor that ITER and planned power plants have chosen (see 2023 below), and helium is the ash of deuterium-tritium fusion, so the two will meet constantly. Exposing polished tungsten to helium plasma in the PISCES-B device, M. J. Baldwin and R. P. Doerner of the University of California San Diego found the surface grew a layer of tangled nanometer-scale rods, more than 5 µm (micrometers) thick after the longest run of about six hours. Working with the NAGDIS-I divertor simulator in Japan, Shin Kajita and colleagues reported the same structure in 2009 and found it formed when the metal was between 1,000 and 2,000 K and the helium ions arrived with more than 20 eV of energy.
Credit
M. J. Baldwin and R. P. Doerner; S. Kajita, N. Ohno and colleagues
2008
Steel tested to 70 dpa
EUROFER97 is Europe's reference steel for the structures behind the wall: a 9% chromium steel with tungsten, vanadium and tantalum, made as a reduced-activation grade, meaning its ingredients were chosen so it becomes less radioactive under neutrons than ordinary steel. Industrial batches of 3.5 and 8.0 tonnes were made, and samples went through irradiation programs in fission reactors up to 70 dpa (displacements per atom: on average, every atom knocked out of its place 70 times) in Russia's BOR-60 reactor. The team found that hardening and brittleness at irradiation temperatures below about 350 to 400 °C did not level off up to 30 dpa and was the limiting factor for the steel. At 70 dpa and 330 °C the yield stress and the ductile-to-brittle transition temperature (DBTT, the temperature below which the steel snaps instead of bending) showed signs of leveling off.
Lessonfor this steel the main danger was not high dose but running cool, below about 350 °C.
Credit
E. Gaganidze and colleagues at Forschungszentrum Karlsruhe, with RIAR in Dimitrovgrad, Fusion for Energy, NRG Petten and SCK-CEN
2013
Metal wall cuts fuel retention
JET replaced its carbon tiles with the materials then planned for ITER: beryllium on the main wall and tungsten in the divertor, the floor of the vessel where the exhaust strikes. Measuring how much gas went in and how much came back out over series of identical pulses, the team found that the long-term rate at which the wall kept fuel had fallen by more than ten times compared with the carbon wall. In high-confinement (H-mode) pulses with 12 MW of heating, for example, it fell from 1.97 × 10^21 to 1.6 × 10^20 deuterium atoms per second.
Credit
S. Brezinsek, T. Loarer, V. Philipps and JET colleagues
2023
ITER drops beryllium
ITER decided in 2023 to armor its first wall, the panels that face the plasma, with tungsten instead of the beryllium it had planned, and made the change part of the new baseline it announced in July 2024. It gave two reasons: the burning-plasma machines planned after it all intend to use tungsten, and beryllium raised problems of dust from erosion, toxicity during maintenance, and modeling suggesting that melting during off-normal events could make panels fail early. The switch means new panels must be developed and bought, but no full first-wall replacement has to be planned, which ITER says saves time and radioactive waste.
Lessona wall material picked for one machine has to match what later machines will use, or what is learned on it does not carry over.
Credit
ITER Organization
2025
Superconducting tape goes industrial
The strongest compact fusion magnets are wound from REBCO tape: a thin film of a ceramic made of a rare-earth element, barium, copper and oxygen, laid on metal tape, which carries current with no resistance when cold. A 2025 review in the journal Superconductivity reported that demand from compact fusion had turned it into an industrial product. The top makers using pulsed laser deposition (a laser vaporizes the ceramic onto the moving tape) each make more than 3,000 km a year, counted as 12 mm-wide tape, and together supply over half of world output, at about $20 per meter. Their tape carries more than 200 A in a 4 mm width at 20 K (−253 °C) in a 20 T field.
Credit
tape makers including Faraday Factory Japan, Shanghai Superconductor Technology and Fujikura; review by Yue Zhao, Yue Wu, Amit Kumar Goyal, H. Huhtinen, P. Paturi and Yuji Tsuchiya
Watch for
A neutron source for materials
IFMIF-DONES, a particle-accelerator facility being built at Escúzar near Granada in Spain, is meant to test materials for future fusion reactors under neutrons; until it runs, fission reactors are the stand-in. Fusion for Energy (the EU's fusion agency), Spain, Japan, Croatia and Italy signed the agreement to build it on 21 November 2025, with Spain paying 55%. The local council granted the building permit on 16 June 2026, with site work scheduled from August 2026, and in October 2026 Fusion for Energy signed its first two procurement arrangements, covering the accelerator's injector and about 15 tonnes of EUROFER97 steel; the main building is expected to be substantially complete by 2029.
Achieved whenthe facility irradiates its first test samples and an outside laboratory publishes the measured neutron flux and damage rate.
Claimed by
Fusion for Energy and the IFMIF-DONES Spain consortium, November 2025 to October 2026
AI-built irradiation database
FusionMatDB, posted on Hugging Face in April 2026, says it is the first publicly accessible, machine-learning-ready database of how fusion materials change under irradiation. It holds 22,269 records with 54 fields each, extracted by a vision-capable AI model from 65 semiannual progress reports of the Fusion Materials Program at Oak Ridge National Laboratory (1990 to 2024) and from ITER's structural design materials library. It notes that EUROfusion's own database, about 3,000 records, is open only to EU consortium members, and it scores its own extraction, with 85% of records at a confidence of 0.7 or more.
Achieved whenan independent check of a random sample against the source reports publishes its error rate (see Open).
Claimed by
the dataset's maintainer, Hugging Face user Khalizo, April 2026
Open
01
Helium tests for AI potentials
A public benchmark that scores ready-made universal machine-learned interatomic potentials, AI models trained to predict the forces between atoms nearly as accurately as quantum-mechanical calculation and far faster, on helium and hydrogen in tungsten: how strongly helium binds to vacancies and to other helium atoms, how hydrogen is trapped, and how helium clusters grow. Each value is compared with density functional theory (DFT), the standard quantum-mechanical method. The best model is then run in a molecular-dynamics simulation to see whether it reproduces the temperatures and ion energies at which tungsten fuzz forms in experiments.
The briefCredit · unclaimed02
Check the AI-built database
An independent audit of FusionMatDB, then an extension. Draw 500 records at random, find each value on the page of the Oak Ridge report it came from, and record field by field whether it matches. Publish the error rate per field and per material class, send corrections to the maintainer, then add open sources the database does not yet cover, each new row linked to its page.
The briefCredit · unclaimed
Open 01 · the brief
Helium tests for AI potentials
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public benchmark that scores ready-made universal machine-learned interatomic potentials, AI models trained to predict the forces between atoms nearly as accurately as quantum-mechanical calculation and far faster, on helium and hydrogen in tungsten: how strongly helium binds to vacancies and to other helium atoms, how hydrogen is trapped, and how helium clusters grow. Each value is compared with density functional theory (DFT), the standard quantum-mechanical method. The best model is then run in a molecular-dynamics simulation to see whether it reproduces the temperatures and ion energies at which tungsten fuzz forms in experiments.
Why it is still open
A 2025 study by Fei Shuang, Poulumi Dey and colleagues found that universal potentials match DFT on many defects in metals and on hydrogen in alloys, but also that data on pure tungsten alone was not enough for complex tungsten defects; a text search of the paper finds no mention of helium. Tungsten comparisons on GitHub cover single properties, such as one vacancy or screw dislocations.
Built with what exists
Open universal potentials such as MACE-MP-0 and the EquiformerV2 models trained on the OMat24 data set, the best performers in that study; the Atomic Simulation Environment (ASE) and LAMMPS to run simulations; Quantum ESPRESSO for new DFT reference values; one rented GPU.
First step
Collect 30 published DFT values for helium and hydrogen in tungsten into a cited table, then score MACE-MP-0 on the first ten in ASE.
Done looks like
At least 30 helium and hydrogen properties in tungsten, each with a cited or newly computed DFT reference; at least five universal potentials scored, with the error for each property; one molecular-dynamics run compared with the experimental window of 1,000 to 2,000 K and ion energies above 20 eV; all inputs and scripts public.
Open 02 · the brief
Check the AI-built database
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An independent audit of FusionMatDB, then an extension. Draw 500 records at random, find each value on the page of the Oak Ridge report it came from, and record field by field whether it matches. Publish the error rate per field and per material class, send corrections to the maintainer, then add open sources the database does not yet cover, each new row linked to its page.
Why it is still open
FusionMatDB reports only its own confidence scores, and a November 2026 search found no outside check of its accuracy and no other open database of its kind.
Built with what exists
The FusionMatDB files and the Oak Ridge reports it cites; a second, independent extraction by a vision-capable AI model working as an agent (see the Note "The jobs agents repeat"); two people reconciling disagreements in a spreadsheet; for the extension, open papers such as the IAEA Fusion Energy Conference paper on EUROFER97 cited above.
First step
Download the data, draw 50 random EUROFER97 and F82H records, and check each against its source page.
Done looks like
500 records checked by two people, with the error rate per field and its 95% confidence interval published; at least 1,000 new records from sources outside Oak Ridge, each with a page-level citation, and an audited sample of 100 of them.