Final disposal of nuclear waste
Spent nuclear fuel must stay isolated for 100,000+ years; the first deep repository for it is only now nearing its operating licence, in Finland.
Open in the interactive tree →Spent fuel stays hazardous for hundreds of thousands of years, so the accepted plan is burial in stable rock several hundred metres down behind several barriers, such as iron-copper canisters in bentonite clay at Onkalo in Finland. The science is largely settled; the hard parts are siting with local consent, licensing, cost and proving safety over timescales no experiment can run. Meanwhile most spent fuel sits in interim storage.
As of October 2026
The IAEA's tool counts 447,758 tonnes of heavy metal of spent fuel discharged worldwide by 24 June 2026, about 322,000 in storage and about 126,000 reprocessed. In Finland the regulator STUK gave Onkalo a favourable safety assessment in August 2026, covering at least 100,000 years; the government's licence decision is expected in autumn and operator Posiva aims to be ready to start disposal by the end of 2026, with drilling of the first deposition hole under way. Sweden began construction at Forsmark on 15 January 2025 for waste in the 2030s; France's inquiry commission gave Cigéo a favourable opinion in August 2026 but the authorisation decree is still missing; Germany's agency plans to hand over proposed siting regions by the end of 2027 after the legal 2031 target was judged unreachable; Switzerland's framework licence procedure for Nördlich Lägern is ongoing. The US has no permanent repository: Congress ended Yucca Mountain funding in 2011.
What is missing
- A first working repository for commercial spent fuel: Onkalo still awaits its government licence and first canister
- Host sites with local consent in countries still searching; on 7 September 2026 Switzerland's federal councillor made fair compensation a condition for the licence
- Cost control: Cigéo's industrial reference cost is 33.36 billion euros (2025 prices), up from 25 billion in 2016
- Evidence that barriers hold for 100,000 years and more, from models, natural analogues and underground laboratories
- A long-term answer for the roughly 322,000 tonnes of spent fuel now in storage
Becomes possible once solved
- Nuclear expansion with a credible answer to the waste question
- Emptying and closing interim storage pools and cask yards
- A template for countries starting nuclear programmes
Open steps
- Better waste glass and ceramics High AI leverageDesign glass and ceramic waste forms that hold more waste and resist dissolution, using machine-learned property models and active learning.
- Fast models of radionuclide transport High AI leverageBuild validated surrogates of coupled flow, chemistry and heat in repository rock and clay, so safety cases can test millions of scenarios over up to a million years.
- AI review of repository safety cases Medium AI leverageTest whether AI can find, cite and cross-check evidence across decades of safety-assessment documents, scored against expert analyses of Yucca Mountain.
- Proving barriers last 100,000 years Medium AI leverageCombine natural analogues, long experiments and models to show that canisters, bentonite buffer and host rock keep radionuclides in for 100,000 years and more.
- Siting with community consent Low AI leverageFind repository sites that meet geology criteria and win local consent, as Germany's search (siting-region proposal due end 2027) and Switzerland's compensation question show.
Where AI could help
Low AI leverage. Consent, licensing, cost and construction set the pace, not computation; AI helps with waste-form design, transport models and safety-case review.
- Design waste glass and ceramics that hold more waste and resist dissolution
- Speed repository transport and barrier simulations with surrogate models
- Search and cross-check decades of safety-case documents
- Optimise canister logistics and repository layouts
Shown so far
- On 22 September 2026, a PNNL-led paper in the Journal of Non-Crystalline Solids reported the first experimental validation of machine-learning-designed low-activity waste glass: retrained models gave better predictions, fewer failures and higher waste loading. source
- On 17 September 2026, the University of Nevada, Reno announced a DOE Genesis Mission project that tests over nine months whether AI can reason from Yucca Mountain safety-assessment evidence as well as established expert analyses; there are no results yet. source
- In February 2026, Germany's MALEK project (about 1.7 million euros, 36 months) was announced to train surrogate models on detailed simulation data for sorption and reactive transport in crystalline rock, with scales from millimetres to kilometres and up to one million years. source
Prerequisites
- Geology and deep time1788
- Nuclear Physics1938
- Nuclear Power1954
- Nuclear Accidents & Safety Turn1979-1986Distrust after the accidents is a main hurdle for siting repositories
- New Reactors & SMRs2020–2026