BRENNEN

Industries / Frontier science

Fusion energy

Fusion energy is the industry that builds machines to heat and squeeze light atoms, usually the heavy forms of hydrogen, until their nuclei fuse, and aims to turn the energy released into electricity.

1951

On 24 March 1951 Argentina's president, Juan Perón, announced that Ronald Richter's Huemul Project had released energy by fusing hydrogen. The claim was false, but it prompted Lyman Spitzer at Princeton to propose the stellarator, a magnetic bottle for hot gas, and the US Atomic Energy Commission funded his Project Matterhorn, where Princeton's fusion research began that year.

iter.orgpppl.gov

39

Milestones achieved across 6 domains.

15

To watch for: announced, not yet confirmed outside.

12

Open deliverables. Nobody has shipped them.

In one minute

Where it started

A false claim from Argentina in 1951 prompted Lyman Spitzer at Princeton to propose the stellarator, and magnetic fusion research stayed secret until 1958, when it was declassified so that all nations could share results. In 1969 a British team measured temperatures of millions of degrees inside the Soviet T-3 tokamak, a doughnut-shaped magnetic vessel, and in 1972 Livermore physicists proposed the other main route: crushing tiny fuel pellets with lasers.

Where it stands

In December 2022 the National Ignition Facility in California got more fusion energy out of a fuel capsule than its laser put in, and in 2023 the JET tokamak in England released a record 69 MJ in one pulse, but no fusion machine has yet made electricity. The International Atomic Energy Agency (IAEA) counted 174 fusion facilities operating, being built or planned in 2026, and describes 26 power plants in planning, eight of them aiming to make electricity by the early 2030s. The Fusion Industry Association says 56 companies it surveyed raised $4.48 billion in the year to July 2026, and ITER, the international machine in France, now plans to start research operation in 2034.

What is open

Much of the open work is checking: tracking whether press-release records are later published, filing forecasts for machines such as SPARC before they run, and running published plant designs, drivers and fuel cycles through one open method so they can be compared. The rest is computing and data work that AI models and agents now put within reach of a small team, such as testing AI models of atoms on tungsten walls, auditing a materials database that an AI model built, and setting a robot benchmark for in-vessel maintenance. None of it needs radioactive material, lasers, strong magnets or a license: every deliverable runs on public data, open-source code or harmless stand-in parts.

Domains

AchievedWatch forOpen

Advances to build with

5 research notes

Open in fusion energy

3 of 12 open deliverables

Open · Magnetic confinement

SPARC forecast, filed in advance

Before SPARC first runs in 2027, gather every published prediction of its performance (fusion gain Q and fusion power, scenario by scenario) into one public table. Add an independent prediction made with an open-source code from CFS's published machine parameters, and archive the whole forecast with a timestamp. After SPARC's first campaign, score each prediction against the measured results.

Built with what exists

TORAX (open source, written in JAX); the machine parameters in the published SPARC papers; Zenodo or a similar archive that issues a dated DOI; a coding agent such as Claude Code to set up the runs and parameter scans, with every input checked by a person.

Credit · unclaimed

Open · Laser and inertial fusion

Shell flaws, an open benchmark

A public, labeled set of microscope images of hollow plastic or glass shells 1 to 4 mm across, the size of inertial fusion capsules, bought commercially and holding no fuel, each photographed from several sides with every surface flaw marked by two people. Add a baseline flaw detector trained on the set, and report how often it agrees with the people compared with how often they agree with each other.

Built with what exists

A stereo microscope with a camera and a motorized rotation stage; an open labeling tool such as CVAT or Label Studio; an open vision model fine-tuned on one GPU (see the Note "One GPU to fine-tune"); Zenodo or Hugging Face to publish the set.

Credit · unclaimed

Open · Materials

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.

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.

Credit · unclaimed