LUNAROPS · OPERATIONAL UPLINK 100% UPTIME 1,247d POSTS 893 JEFF.MOON@LUNAROPS.DEV UTC --:--:--

How Nuclear Reactors Work

16 min read Infrastructure of Civilization
nuclear-powerenergyfissionpower-gridcontrol-systemsinfrastructuresafety-engineering
Contents

A nuclear reactor is a machine for holding a number at exactly 1.000. The number is the neutron multiplication factor — how many neutrons from this generation of fission events survive to cause the next generation — and everything about reactor design is in service of keeping it there: the geometry of the core, the chemistry of the coolant, the metallurgy of the fuel, the control systems, and several layers of physics that push the number back toward 1 when it drifts. What makes reactors interesting to an engineer is not that they’re exotic; it’s that they are the most consequential feedback-control problem ever productized, with thermal margins measured in minutes and consequences measured in decades. And the punchline of the whole field, which this post will earn by the end: the physics was essentially solved by 1960, the safety engineering by 2000, and what actually determines whether nuclear power grows or dies is interest rates on construction loans.

This is the first post in the Big Energy series, and a companion to How the Power Grid Works — reactors are the grid’s largest synchronous machines, and their inertia matters there. Here we go inside the containment building.

Fission, and the Accident of Controllability

When a neutron is absorbed by a uranium-235 nucleus, the nucleus deforms, splits into two smaller nuclei, and releases about 200 MeV of energy — roughly 80 million times the energy of breaking a chemical bond — plus, critically, two or three fresh neutrons. Those neutrons can cause further fissions, which release further neutrons. That’s the chain reaction, and the bookkeeping is one number:

k_eff = neutrons causing fission in generation N+1
        ------------------------------------------
        neutrons causing fission in generation N

k < 1   subcritical  — reaction dies out
k = 1   critical     — steady state ("critical" means STABLE, not "about to explode")
k > 1   supercritical — power rising

Energy bookkeeping:
  1 fission of U-235      ~200 MeV  ~= 3.2e-11 J
  1 kg of U-235 fissioned ~= 8.2e13 J ~= 950 MWd of heat
                          ~= the energy in ~3,000 tonnes of coal

A reactor at full power and a reactor at 1% power are both exactly critical; the power level is just where you parked the reaction when you brought k back to 1.

Here is the fact the entire industry rests on, and it is a genuine accident of nuclear physics. Most fission neutrons — about 99.35% of them — are “prompt”: they appear within about 10 microseconds. If those were the only neutrons, the time between generations would be around 100 microseconds, and any tiny excess of k over 1 would compound through ten thousand generations per second. Power would double in milliseconds. No control rod, no pump, no human, no computer moves fast enough; the machine would be uncontrollable in principle, not just in practice.

But 0.65% of neutrons (for U-235; the fraction is called beta) are “delayed” — they come not from fission itself but from the radioactive decay of fission fragments, seconds to minutes later. As long as the reactor needs that delayed fraction to stay supercritical, the effective generation time is dominated by those slow precursors, and power changes on a timescale of tens of seconds. That’s a control loop a 1950s analog system — or a human with a lever — can close comfortably.

Reactivity rho = (k - 1) / k        (excess above critical)

  rho < beta (0.0065):  delayed-neutron controlled — period of seconds
                        to minutes. Normal operation lives here.
  rho > beta:           "prompt critical" — prompt neutrons alone
                        sustain the reaction. Period of milliseconds.
                        This is the line you never cross.

Chernobyl crossed it. Every reactor design decision since has been, in part, about making that line unreachable.

Moderation: Why Water Is Both Coolant and Brake

Fission neutrons are born fast — around 2 MeV, moving at ~7% of light speed. But U-235’s appetite for neutrons is hundreds of times larger for slow (“thermal”) neutrons, around 0.025 eV, ambling along at thermal-equilibrium speeds. So most reactors slow their neutrons down using a moderator: a material full of light nuclei that neutrons can dump energy into through collisions, like a cue ball hitting balls of similar mass.

Ordinary water is a superb moderator — hydrogen is the lightest nucleus there is — and it’s also a superb coolant, which is why about 85% of the world’s power reactors are light-water reactors (LWRs) that use the same water for both jobs. That double duty is not a cost compromise. It is the central safety feature of the architecture:

If the water goes away, so does the chain reaction. Boil the water into steam (much less dense), or lose it entirely, and the neutrons stop being moderated, stay fast, and stop fissioning U-235 efficiently. The reaction shuts itself down. This is a negative void coefficient: voids (steam bubbles) in the coolant reduce reactivity. The laws of physics, not a safety system, are the first responder.

The same logic operates continuously as temperature feedback. Hotter water is less dense, moderates less, and pushes k down. Hotter fuel broadens the U-238 absorption resonances (Doppler broadening), so more neutrons get captured uselessly, pushing k down again — and this one acts in the fuel itself, instantly. An LWR that drifts hot pulls itself back; an LWR that drifts cool pushes itself up. The reactor is a self-regulating system with a stable setpoint, and the operators are mostly trimming it.

This is also exactly what the RBMK — Chernobyl’s design — got wrong. It used graphite as the moderator and water only as coolant. Boil the water and you lose neutron absorption but keep the graphite moderation: reactivity goes up. A positive void coefficient, a reactor that responds to overheating by making more heat. Combined with a control-rod design flaw and a test run in a forbidden regime, that feedback sign error is the physics core of the worst accident in the industry’s history. No Western power reactor has ever had it, and no reactor licensed today is allowed to.

The Machine Itself: PWR vs BWR

Two architectures dominate. Both are LWRs; they differ in where the steam is made.

PRESSURIZED WATER REACTOR (PWR) — two loops

   CONTAINMENT                          |  TURBINE HALL
                                        |
   +--------------+    ~330 C, 155 bar |
   | control rods |   (kept liquid)     |
   |   |  |  |    |  primary loop       |   secondary loop (steam)
   |   v  v  v    |  +--------------+   |   +---------+    +-----------+
   |  [ CORE ]----+->| STEAM        |---+-->| TURBINE |--->| GENERATOR |
   |  fuel rods   |  | GENERATOR    |   |   +----+----+    +-----------+
   |              |<-| (heat xfer,  |<--+--------+ condenser,
   +--------------+  | no mixing)   |   |          cooling tower
        ^            +--------------+   |
        |  pressurizer keeps primary    |  radioactive water never
        |  water liquid at 155 bar      |  leaves containment

A PWR keeps its primary water under ~155 bar of pressure so it stays liquid at 330°C, and transfers heat to a separate secondary loop in steam generators. The turbine sees clean, never-irradiated steam. Two loops cost money and the steam generators are a maintenance headache (tube degradation has killed more than one plant’s economics), but the radioactive inventory stays inside containment and the core chemistry can be tuned independently — including dissolving boron (a neutron absorber) in the primary water as a slow control knob.

A BWR deletes the steam generators and boils water directly in the core at ~70 bar; the steam drives the turbine directly. Simpler, fewer components, lower pressure vessel stress — but now the turbine hall is plumbed to the reactor core, mildly radioactive during operation, and control is partly done by changing recirculation flow (more flow sweeps bubbles out, adds moderation, raises power — the void coefficient working as a throttle). Fukushima Daiichi’s units were BWRs, though what failed there was not the BWR concept, as we’ll see.

Type Moderator Coolant Fuel World fleet share Defining trait
PWR Light water Light water, 155 bar liquid UO2, 3–5% enriched ~70% Two loops; boron chemistry; clean secondary side
BWR Light water Light water, boils in core UO2, 3–5% enriched ~14% Direct cycle; simpler; radioactive turbine side
PHWR / CANDU Heavy water Heavy water Natural uranium (0.7%) ~10% No enrichment needed; refuels online
RBMK Graphite Light water ~2% enriched <2%, legacy only Positive void coefficient; Chernobyl’s design
AGR (UK) Graphite CO2 gas ~3% enriched legacy only High thermal efficiency, niche

The CANDU deserves its parenthetical respect: by moderating with heavy water (deuterium absorbs almost no neutrons), it runs on natural uranium with no enrichment plant anywhere in its supply chain, and swaps fuel bundles while running at full power — a 90%+ capacity factor design from a country that decided in 1950 not to build enrichment infrastructure and engineered around it.

Control in all of these is layered like any good system: control rods (boron carbide or silver-indium-cadmium absorbers, driven into the core in seconds for a scram, milliseconds-to-insert by gravity in a PWR), soluble boron for slow trim over a fuel cycle, burnable poisons baked into fresh fuel that fade as the fuel does, and the inherent temperature feedbacks underneath everything. One operational wrinkle worth knowing because it shaped history: the fission product xenon-135 is the most voracious neutron absorber known, builds up after a power reduction, and can make a recently-downpowered reactor temporarily impossible to restart — the “xenon pit.” Impatience with a xenon-poisoned core, and rod withdrawals to fight it, is how Chernobyl’s crew got into their forbidden regime.

Decay Heat: Why You Can’t Just Turn It Off

Here is the single most important engineering fact about nuclear power, and the one the public conversation misses almost entirely. When you scram a reactor, the chain reaction stops in seconds. The heat does not.

The core is full of fission fragments — hundreds of isotope species, most of them radioactive — and their decay deposits heat no control rod can touch. It follows a stubborn power law:

Decay heat after shutdown (typical large core, 3,000 MW thermal):

  t = 0        ~6.5%  of full power   ~200 MW   (a small power station)
  t = 1 hour   ~1.5%                  ~45 MW
  t = 1 day    ~0.5%                  ~15 MW
  t = 1 week   ~0.25%                 ~8 MW
  t = 1 year   ~0.05%                 ~1.5 MW

You cannot switch this off. You can only cool it.

Every nuclear accident that has ever released significant radiation — Three Mile Island, Chernobyl’s aftermath, Fukushima — is at root a decay heat removal failure, not a runaway chain reaction (Chernobyl uniquely had both). Fukushima is the cleanest case study: all reactors scrammed successfully on the earthquake, exactly as designed. The chain reactions were dead forty minutes before the tsunami arrived. What the wave killed was the diesel generators, and with them the pumps — and 1.5% of 1,500–2,400 MWt, with nowhere to go, boiled away the coolant, overheated the fuel cladding until its zirconium stripped oxygen from steam (producing the hydrogen that blew the roofs off), and melted three cores over the following days. The lesson written in that accident: a reactor’s safety case is only as good as its ability to reject a few tens of megawatts of heat for weeks, with no grid, no diesels, and no heroics. This is the “station blackout” scenario, and it — not bombs, not meltdowns-from-runaway — is what modern reactor safety engineering actually targets. The post-incident reviews that reshaped the industry worldwide are, incidentally, a masterclass in the discipline covered in The Art of the Postmortem.

It’s the same problem every thermal engineer knows — your rack is also a device that converts all input power to heat, per the thermodynamics of cooling — except here the load shedding option doesn’t exist. The heat comes whether you want it or not.

Gen III+: Safety by Gravity

The generation of reactors designed after these lessons — Gen III+ — made one decisive move: remove the dependence on powered equipment for the decay-heat problem. The flagship is Westinghouse’s AP1000 (the units now running at Vogtle in Georgia and at several Chinese sites).

In an AP1000 station blackout, no diesels, no operator action:

  • A huge water tank sits above the core inside containment; valves held shut by powered solenoids fail open on power loss, and water falls into the core by gravity.
  • The steel containment vessel itself becomes the heat exchanger: decay heat boils water, steam condenses on the cool steel shell, runs down the walls, and returns to the core. Outside, air convects up around the shell in a natural chimney, with another gravity-drained tank wetting the exterior for evaporative cooling.
  • The plant rides this out for 72 hours with zero AC power and zero human intervention, and indefinitely if someone can refill a tank on the roof with a fire truck.

The physics doing the work is density differences and gravity — things with five nines of uptime. Europe’s EPR took the complementary belt-and-suspenders path: quadruple-redundant active safety trains plus a “core catcher,” a ceramic spreading basin under the vessel that assumes the worst has already happened and is designed to cool a molten core where it lands. Both philosophies are honest engineering; the AP1000’s is cheaper to build and the EPR’s has proven brutally expensive (Flamanville and Olkiluoto each ran more than a decade late).

It’s worth stating plainly, because the record supports it: light-water nuclear power, measured in deaths per terawatt-hour including its accidents, sits alongside wind and solar and orders of magnitude below every fossil fuel. The safety problem is, by the standards of energy infrastructure, solved. That is not the industry’s problem.

The Fuel Cycle, Briefly

Natural uranium is 0.72% U-235; LWR fuel needs 3–5%. Enrichment is done in cascades of gas centrifuges spinning uranium hexafluoride, and the work is measured in Separative Work Units (SWU) — the same machines and math that make 90% weapons-grade material, differing only in how long you keep feeding the cascade, which is why enrichment is the proliferation chokepoint and why the international monitoring regime obsesses over it. (Defense of industrial monitoring systems generally being its own discipline — see the SCADA tour.)

The enriched UO2 is sintered into ceramic pellets, stacked in zirconium-alloy tubes, and bundled into assemblies; a big PWR holds ~150–200 of them and swaps out a third every 18–24 months, each batch having delivered a “burnup” of 45–60 gigawatt-days per tonne. Spent assemblies are ferociously radioactive and thermally hot — back to decay heat — so they cool in a deep pool for 5–10 years, then move to passively air-cooled concrete-and-steel dry casks on a pad behind the plant. Dry casks are boring, safe, and a genuinely adequate solution for a century; the political failure to site a permanent geological repository (the US has collected ~$40B in ratepayer fees for one and built nothing usable) is a policy embarrassment, not an engineering gap. France reprocesses spent fuel to recycle plutonium into fresh fuel; it works, reduces waste volume, and has never been economic at prevailing uranium prices — the once-through cycle persists because uranium is cheap, not because recycling is hard.

SMRs, Honestly

Small Modular Reactors are the industry’s bet that its disease is bespoke megaprojects, and the cure is factory production: reactors of 50–300 MWe built on assembly lines, shipped to site, installed in series. The physics is fine — most designs are conservative small LWRs, and decay heat removal gets easier as cores shrink (surface-to-volume ratio works for you; many designs can passively cool forever).

The honest scorecard, though, is mixed. NuScale, the US frontrunner, got the first SMR design certification through the NRC — then watched its launch project (UAMPS, Utah) collapse in late 2023 when the target power price rose from $58 to $89/MWh before a single module was fabricated, and the customers walked. The cancellation wasn’t about safety or physics; the economies of scale that SMRs surrender are real and immediate, while the economies of series they promise only arrive after dozens of units — a chicken-and-egg problem in financing. Meanwhile Russia operates a floating twin-SMR plant and China has grid-connected its HTR-PM and is building Linglong One; the technology demonstrably works. The open question is purely whether anyone reaches the production volumes where the factory thesis pays. The most plausible path may be the least romantic: single-purpose fleets bought by deep-pocketed customers with desperate load growth — which is why the hyperscalers’ datacenter power deals (and GE-Hitachi’s BWRX-300 orders) are the SMR news actually worth watching.

The Real Constraint Is a Spreadsheet

Here is where the series thesis lands. A nuclear plant is the cheapest electricity you can buy from the day it’s paid off — fuel is ~5% of its cost structure, capacity factors run 90%+, and plants are now licensed to 80 years. Vogtle’s new units, the poster children for everything wrong with Western nuclear construction ($35B for two units against a $14B estimate, seven years late), will still likely produce power into the 2090s.

The problem is everything before that day:

Why construction cost dominates:

  LCOE ~= (capital recovery + O&M + fuel) / MWh produced

  For nuclear: capital ~ 70%+ of LCOE. So LCOE is hostage to:
    - overnight cost ($/kW): 2-3x higher in US/EU than China/Korea
    - construction time:  every year of delay compounds interest
                          on billions already spent, producing nothing
    - cost of capital:    at 3% vs 9% discount rates, the SAME plant
                          roughly doubles its lifetime cost of power

  Nuclear is less an energy technology than a financing problem
  wearing a containment dome.

The proof that the costs aren’t intrinsic is that other builders don’t pay them. France built 56 reactors in about two decades by building the same design repeatedly with a standing workforce; South Korea and China do the same today at a third of Western unit costs and in 5–6 years per unit. The US and Europe, by building one-of-a-kind plants once a decade with workforces assembled from scratch and regulatory redesign mid-construction, sit at the wrong end of the learning curve and pay the wrong end’s prices. Every credible path to cheaper Western nuclear — SMR factories, standardized Gen III+ fleets, build-the-same-thing-again at existing sites (Vogtle 4 came in ~30% cheaper than Vogtle 3 for exactly this reason) — is a variation of one idea: stop building first-of-a-kind machines.

Verdict

Understand a reactor as three nested systems and the whole field falls into place. Innermost: a chain reaction made controllable by a 0.65% accident of physics — delayed neutrons — and made self-stabilizing by negative temperature and void feedback, with water as both working fluid and failsafe. Middle: a decay-heat removal machine, because the one thing fission can’t do is stop making heat on command — this is what containment, diesels, gravity tanks, and the whole Gen III+ passive-safety movement actually exist for, and what failed at Fukushima. Outermost: a thirty-year financing instrument, where interest rates and first-of-a-kind construction risk — not physics, not safety, not waste — decide whether anything gets built.

For an engineer reading the energy news: discount any nuclear claim that talks physics breakthroughs, and weight heavily any news about standardization, repeat builds, and cost of capital. The reactor was never the hard part. The construction site is.

Sources

Comments