How to fuse ordinary hydrogen
Ordinary hydrogen (protium) is abundant, and stars fuse it into helium. However, protium fusion is monumentally harder than current fusion research, which instead uses the rare isotopes deuterium and tritium. To my knowledge, no one has sketched out a design for a future practical protium fusion reactor. This is surprising to me because people speculate in great technical detail about how spacefaring civilizations could move stars around, disassemble them, and capture their output to run huge computers. There is even a famous proposal called the Bussard ramjet to use protium fusion for interstellar travel, but this doesn't pencil by several orders of magnitude.
So, will a protium fusion reactor that isn't just a star ever be practical?
I looked into this question with substantial help from GPT-6 Astra and 6.1 Terra, and it seems the answer is yes. This page is a design sketch for a reactor buildable by a far-future civilization. Briefly, it uses an enormous ~2 million km wide reactor, using a CNO cycle, pulsed inertial confinement, and pulses large enough (~10 billion tons each) to be optically thick.
This design is limited in several ways. First, it's not clear if all the details will actually work out (I'm about 80% sure something like this design can work, but maybe you should take this with a grain of salt because I'm not a plasma physicist, just some guy with a Codex subscription.) Second, it can't be used for interstellar propulsion like the Bussard ramjet. Finally, it has low power density, only about 278 mW/kg, which is much worse than solar panels, fission or other fusion reactors.
However, this minuscule power density is still 1,500x better than the Sun's 0.19 mW/kg, meaning that with reactors the mass of the Sun, we can utilize the Sun's energy 1,500x faster than with a Dyson sphere.[1] There is also substantial slack in the design (I may explain this later, but for now here's a footnote from GPT-6.1)[✨2], such that I think a mature civilization will actually achieve at least 1 W/kg, and possibly more than the 6-12 W/kg Eddington limit possible for any star.
Why is protium fusion so hard?
Most fusion research uses deuterium and tritium because they fuse much more readily than ordinary hydrogen. At a collision energy of 10 keV, a typical energy for fusion research, the proton–proton reaction has a cross section about 1024 times smaller than deuterium–tritium (D-T) fusion. Part of the reason is that one of the protons must turn into a neutron through the weak nuclear force as they fuse.[✨3] The Sun can make this work because it has an immense amount of hydrogen and billions of years. A reactor has a lot less time.
This is a problem for the Bussard ramjet, which Robert Bussard proposed in 1960 to collect interstellar hydrogen and fuse it to propel a spacecraft. In 1975, physicist Daniel Whitmire calculated that proton–proton fusion would be far too slow to power a Bussard ramjet. His calculations imply that a reactor with a 19-meter-radius burning region would produce about 10¹⁷ times too little power to accelerate a 1,000-ton spacecraft at 1 g.[✨4]
There is another way to fuse hydrogen: the carbon–nitrogen–oxygen (CNO) cycle, which dominates in stars larger than the Sun. In the CNO cycle, carbon, nitrogen and oxygen nuclei form a sequence that absorbs four protons, emits one helium nucleus, and returns to the starting nucleus. The catalyst provides a path around the slow proton–proton reaction.[source]
In 1975, Whitmire proposed using CNO and other catalytic cycles to power a Bussard ramjet. For that spacecraft, he estimated that a CNO reactor could produce enough fusion reactions at that size. That solves the reaction-rate problem on paper. But there's another problem: bremsstrahlung, or X-rays emitted by electrons as they pass charged nuclei, which can bleed energy away from the hot plasma.
In a D-T reactor, fusion can gain more power than bremsstrahlung loses. But even with the CNO cycle, the vastly lower reaction rates of protium fusion mean a transparent plasma is unable to keep up with bremsstrahlung. On top of this, CNO's higher temperature and heavy catalyst nuclei both increase bremsstrahlung, which scales roughly as (electron temperature)1/2 and (nuclear charge)2. For one hydrogen–nitrogen mixture, electron–ion bremsstrahlung is about 16 times as high as in a D-T plasma.[✨5] A rough estimate for Whitmire's design puts bremsstrahlung losses about 2,000,000 times higher than its fusion power, making it catastrophically unable to sustain its burn. Whitmire recognized the radiation problem and hinted at several solutions, but did not find a design that would resolve it.[✨6]
The way around this explored here is to make the fuel dense and optically thick, as in stars. If X-rays interact many times before escaping, absorption can return their energy to the fuel, and scattering can slow their escape. The radiation still exerts pressure, and it eventually has to be captured, but the fuel may stay hot long enough to fuse. This is why the reactor uses extremely large, dense pulses.
Reactor design
In this design, we imagine that a far-future civilization builds a spherical reactor whose first wall has a radius of about 1.1 million km. The reactor masses 3.56 × 1022 kg, or 0.48 times the mass of the Moon. Launchers accelerate centimeter-scale fuel pellets containing hydrogen and C/N catalyst toward the center, at speeds of up to ~3,700 km/s. Like an implosion-type nuclear bomb, the reactor compresses material by driving it inward from many directions. The pellets’ inward momentum briefly compresses the fuel to fusion conditions (as hot as 1.6 GK), achieving a gain of about Q=7.73-22.66 depending on the catalyst isotope. The hot products expand and deposit their kinetic energy into the graphite first wall. The wall also absorbs prompt radiation and supplies heat to a heat engine, and collection equipment recovers catalyst for later pulses.
Operating sequence
↺ Fuel returns; helium leaves.
The CNO cycle
In this reactor's version of the CNO cycle, there are four stages, each of which corresponds to its own pulse in the reactor. Follow the catalyst through the reactions using the diagram below.
Hover over, focus on, or tap an isotope or pulse stage to explore its role in the cycle.
Over the complete loop, a catalyst nucleus absorbs four protons and emits a helium nucleus. The reactions release about 26.7 MeV per completed helium nucleus. Two beta decay events emit neutrinos that escape with part of the energy, so only 24.7 MeV is available after neutrino losses.
Every cycle, a very small percentage of catalyst is replaced. This is because a small side branch of the nitrogen-15 pulse results in oxygen-16, which the plant removes and replaces with fresh carbon.
Fuel recovery
A collection layer in the graphite inner wall captures products. Warm pumps move recovered hydrogen and helium into a buffer vessel, then to separation equipment. From there, hydrogen and reusable catalyst return to the cycle, while helium is removed. The layer is repaired and periodically processed to recover trapped material. Explore the first wall and fuel-recovery system →
Electricity generation
The first wall supplies heat to a heat engine. The plant uses some of the electricity to prepare later pulses and operate recovery, support and maintenance equipment. Of the nuclear energy released, about 7.6% leaves as neutrinos, 34.7% is exported as electricity and 57.7% is radiated away as waste heat. The plant exports 9.87 × 1021 W from 3.56 × 1022 kg of equipment, or 277.68 mW/kg. See the energy balance →
Economics
At 277.68 mW/kg, the reactor has much lower power per kilogram than solar panels or nuclear reactors. If construction takes 100 MJ per kilogram of plant, it would take about 11.4 years of net electricity output to repay that energy. While this is a rather capital-intensive way to generate energy, the main appeal is that runs on abundant ordinary hydrogen and a small supply of carbon catalyst, making long-lived power possible wherever those are available. See the energy and resource comparisons →
Will this actually work?
Because I know very little actual physics and relied on GPT for essentially the entire design sketch, you might think this is entirely spurious physics and engineering. However, this seems fairly unlikely to me:
- The basic physical principles and reaction rates are established physics. The CNO cycle was hypothesized to exist in the late 1930s, has undergone over 80 years of study, and was in 2020 observed using solar neutrinos. Reaction rates of all these reactions have likewise been experimentally observed, and are fairly well constrained, with the exception of 13N + p → 14O, whose uncertainty is not load-bearing.
- Most of the major components have some real-world precedent. For example, the wall temperature of 1600K requires further assessment of irradiated material lifetime. The accelerating field of the launcher, on the order of 10 MV/m, is similar to that of the LHC's 5 MV/m. To the extent they don't, the assumptions are well within what should be possible under known fundamental physics and more conservative than speculative engineering claims in works like Drexler's Nanosystems. For example, Drexler thinks nanotech can perform nearly any chemical transformation with 99% energy efficiency, but I only assume that key reactor components can be manufactured and refurbished with 50% efficiency.
- Language models are just smarter these days. When I first tried this with GPT-5.5 and Claude Opus 4.8, they would keep making obvious mistakes such as "forgetting" about bremsstrahlung that I would repeatedly correct them about, such that I stopped trusting that they had understood all the relevant considerations. Although GPT-6 Astra and 6.1 Sol have pretty bad research taste, they're good at the physically grounded work required here. Instead I asked the models to spend about one Pro 200 weekly allowance getting to 1 mW/kg, then clarified some overly restrictive assumptions it made, understood what I could of the design, and spent about one more weekly allowance getting to 278 mW/kg.
With that said, here is my best guess at the remaining risk and the path forward.
I asked GPT-6.1 Sol to divide the design into 9 components that could pose technical risk, and also had GPT-6.1 Sol and Terra generate the top remaining questions. Of these, it seems like the launchers, accurate placement of pellets, and graphite first wall plates have large technical risk. I also don't understand to what extent integration of the components will present issues, or if some components can make up for shortfalls in the performance of others. Most of these seem fairly tractable for agents to research autonomously with some oversight from me, and push this from "design sketch" to more of a "preliminary design".
See the design status and open questions →
I would greatly appreciate review by a human physicist who could tell me if I'm completely off base here. But because of the above, plus the substantial margin in the design mentioned above in [✨2], this effort so far seems like significant evidence that practical artificial protium fusion is possible and, if it is economically useful, would be conducted by future spacefaring civilization.
I plan to expand this writeup in three directions in the future. First, develop a more aggressive design in the 1-10 W/kg range. Second, reduce the uncertainties in this 278 mW/kg design by starting from the aggressive design and relaxing assumptions in a sensible manner. Third, look into the industrial relevance, as it's not clear where these fusion reactors would fit in a universe that has Dyson spheres and where civilizations can, for example, use the Penrose process to turn spare matter into energy by throwing it into black holes. My guess is it will take two or three more Pro 200 weekly allowance equivalents to remove 30-50% of the remaining uncertainty one can address without doing physical experiments with current models, or much less with future model generations.
Footnotes
- [1] To do this we'd need enough heavy elements for a fleet of reactors with a combined mass comparable to the Sun's, so it isn't quite a fair comparison. ↩
- [✨2] The design retains substantial room for improvement, with uneven margins across its components. The large nitrogen-15 pulses and 400 K radiators already provide much of the gain in this design. The serialized pulse schedule occupies about 28% of the calendar, but higher throughput also requires faster preparation, collection and heat removal. The radiators use about 83.5% of the outer surface against a 90% allocation limit, leaving about 8% more heat-rejection capacity at the same temperature and emissivity. The equipment study retains 1,000 kg/m² of service equipment; reducing this to 700 kg/m² gives about 331 mW/kg, conditional on compact local loops and a complete equipment bill. Some margins are much tighter: the cold-equipment screen leaves only about 2 W/m² within its electrical allowance, and receiver lifetime and preparation of the enlarged pulses remain to be established. Further gains could come from lighter equipment, better heat engines or direct energy recovery. Reaching 1 W/kg would require roughly 3.6 times the present output per kilogram and a new whole-plant account. ↩ ↩
- [✨3] Two protons do not form a stable bound nucleus through the strong force alone. In the proton–proton reaction, one converts to a neutron through the weak interaction, producing a positron and a neutrino. At 10 keV center-of-mass energy, the proton–proton cross section estimated from the astrophysical S-factor in Solar fusion III is about 3.6 × 10−26 barns, versus about 0.027 barns for deuterium–tritium; see also fusion cross-section measurements and evaluated data. This compares individual reactions at the same collision energy; reactor performance also depends on temperature, density and confinement. ↩
- [✨4] In Whitmire's example, the spherical fusion plasma has a temperature of 1 GK and a proton density of 5 × 1019/cm³. He estimated that a proton–proton burning region with a radius of 7,350 km would produce the required reactions; his CNO example had a radius of 19 meters. Here radius means the distance from the center to the edge of the assumed spherical burning plasma. At the same proton density and temperature, a 19-meter-radius proton–proton sphere would produce only (19 m / 7,350 km)³, or about one part in 6 × 1016, of that reaction rate. This is an order-of-magnitude comparison using Whitmire's reported figures, before radiation losses and engineering are accounted for. ↩
- [✨5] This compares a fully ionized plasma with equal numbers of hydrogen and nitrogen nuclei at an electron temperature of 1 billion K with a pure D-T plasma at 150 million K, both at the same electron density. At those temperatures, a nitrogen nucleus produces about 130 times as much electron–ion bremsstrahlung as a tritium nucleus. For the whole mixture, the nuclear charges add a factor of 6.25 and the temperature adds a factor of √(1 billion / 150 million), giving about 16 times as much. This is a classical estimate; relativistic and other corrections affect the exact result. See the Naval Research Laboratory Plasma Formulary. ↩
- [✨6] The two-million figure is our own rough calculation. Whitmire's paper says radiation losses exceed energy production at this temperature, but does not give the ratio. It assumes equal number densities n of protons and fully ionized nitrogen-14, equal electron and ion temperatures of 1 GK, and freely escaping radiation. Using a classical electron–ion bremsstrahlung approximation with the Gaunt factor set to 1, the radiated power is approximately 1.4 × 10−27 √T ne ΣZi2ni erg/cm³/s, with temperature in kelvin and number densities in cm−3; here ne = 8n and ΣZi2ni = 50n. Using Whitmire's nitrogen–proton rate of 2.1 × 10−22 cm³/s and crediting the entire 26.7 MeV CNO-cycle energy to each nitrogen capture gives radiation about two million times the fusion power. Neutrino losses and other radiation would worsen the balance. The reaction rate is from Whitmire's paper; equation 30 in the Naval Research Laboratory Plasma Formulary gives a comparable bremsstrahlung formula with a slightly different coefficient. ↩
Sources and calculations
- Bussard, Robert W. (1960). “Galactic Matter and Interstellar Flight.” Astronautica Acta 6, no. 4: 179–194.
- Solar fusion III: New data and theory for hydrogen-burning stars — a review of the nuclear reactions and their uncertainties. ↩
- An experimental constraint on the nitrogen-13 proton-capture rate — a transfer-reaction study relevant to the hot CNO cycle.
- Rosseland and flux mean opacities for Compton scattering and thermal bremsstrahlung calculations — relevant radiation physics.
- Formation of a spherical plasma liner — an experimental study of merging discrete plasma jets.
- Multi-dimensional modeling of the Plasma Liner Experiment — examples of linking liner simulations across physical models and dimensions.
- The JINA REACLIB reaction-rate database and measurements of competing nitrogen-15 reaction channels — nuclear data relevant to the network and its side branches.
- TUNL evaluated nuclear half-lives — decay data for the pulse products.
- Geant4 electromagnetic physics — the transport models used for the first wall calculation.
- Graphene membrane experiments and hydrogen permeation through graphene — material mechanisms relevant to the cap, with behavior under irradiation and first wall conditions still to be established.
- Allocated plant and isotope-flow data and catcher-interface data — the pulse and isotope data and the plant balance, respectively. The design repository contains the calculations, supporting research and available data.