Nuclear Reactor Design for Young Inventors

For engineering and physics students/instructors who want a hands-on laboratory.

Comment

Inspiration

I created a nuclear design engineering project many years ago that did back-of-the-envelope calculations. I used that as the inspiration to create a Claude Code (Opus 5) interactive engineering trade-off design team of agents that redesigned and rebuilt that simplistic concept into a sophisticated, fully functioning Web-based interactive engineering/physics application.

Reactor design is one of the few engineering disciplines where the interesting part is not any single calculation — it is the trade-offs that create options. Enlarge the core, and you fix your heat flux but pay in fuel inventory. Pick graphite, and you get superb neutron economy but a core the size of a house. You cannot learn that from a textbook derivation. You learn it by running fifty cases in an afternoon and watching the numbers fight each other.

What it does

Uses AI to create an interactive engineering trade-off design team of agents that builds a Web-Based interactive engineering application.

You choose a fissile fuel from six isotopes, a moderator from five, a coolant from four, and one of three core geometries. You set five operating parameters — maximum and minimum coolant temperature, tube diameter, porosity, and coolant flow rate.

The program then solves for the moderator-to-fuel atom ratio that makes your reactor exactly critical, and from that ratio derives everything else: critical mass, geometric buckling, fuel number density, thermal power, power density, average heat flux, fuel centreline temperature, specific power, thermal neutron flux, annual burnup, coolant velocity, Reynolds number, core pressure drop, pumping power, secondary flow rate, turbine output, and cycle efficiency.

There is no calculate button. Every number recomputes the instant you move a slider, and a live plant schematic redraws itself for your core shape and for open versus closed loop power conversion. You can save cases and compare them side by side, which is how a parametric study actually gets done.

If your design cannot go critical, a banner says so immediately and names the three variables that will fix it.

How we built it

Starting from engineering principles developed years ago, I used Claude Code to create an interactive engineering trade-off design team of agents and checked the results using Gemini.

Every cross section, every material property table, every temperature-keyed coolant step table, and every correlation carried over. The Blasius friction factor, the Fermi age treatment, the buckling formulas for all three geometries: all period-correct.

Then a modern shell around it. Next.js and React for the app, Tailwind for styling, and hand-written SVG for the plant schematic, because the original's charm was its diagram and a charting library would have flattened it. The whole thing is a static site — the solver runs in the browser, no backend, no data leaves the machine.

Then the manual, generated rather than written. Every number quoted in the book was produced by running the actual solver, not typed from memory. manual/cases.mts drives the solver across every worked example and every parametric sweep and prints the results; manual/build.mjs assembles the .docx from those numbers plus nine generated figures. If the physics changes, the book can be rebuilt and stays correct.

Built with Claude Code (Opus 5). The collaboration model that worked was: I supplied the domain judgment about what the program was for, and the AI did the archaeology, the port, the verification, and the document generation — with every physics claim checked back against the original source rather than generated from general knowledge.

Challenges we ran into

Debugging was a problem.

Accomplishments that we're proud of

It runs. The solver produced all of the engineering and physics trade-offs, and a script regenerates the document from the solver's output.

Above that, critical mass falls steeply to a minimum of 4.09 kg at a 48 cm cube, then climbs again — 331 kg at 300 cm, 1509 kg at 500 cm. Two effects run against each other: required fuel concentration falls as leakage relaxes, while volume grows as the cube of the dimension. Below the minimum, concentration wins; above it, volume wins. Nobody plotted that in 1993 because each point took an evening. Now it takes a script.

The Am-242 result is a genuinely good teaching moment. In the same 200 cm cube, U-235 needs 100.8 kg and americium-242 needs 9.2 kg — an eleventh, tracking the absorption cross section almost exactly. It is also completely impractical, and the manual says so. Showing a student why an obviously superior number is useless is worth more than hiding it.

What we learned

The new AI models are becoming more capable of enabling science to be updated and more powerful engineering and physics designs to be resurrected from works of the past.

A tool's most valuable teaching moment can be its own failure. The default case being thermally impossible looked like a flaw. It became Worked Example 1 in the manual — read the warning signs, diagnose, enlarge the core to 27 m³, and watch heat flux fall from 798 to 29.6 W/cm² while critical mass rises from 14.8 kg to 331 kg. That trade is reactor design, and no amount of exposition teaches it like watching it happen.

What's next for Nuclear Reactor Design for Young Inventors

Nuclear reactors are needed for the developing world, and engineers and scientists need to be trained to meet the demand. Web tools such as this could be a valuable educational device.

For this project, we could add a reflector. The single largest gap. Every core is currently bare, so every critical mass is conservative by roughly a third. A one-group reflector saving would make the numbers markedly more realistic.

Make cycle efficiency respond to primary temperature. Right now the secondary state points are fixed, so a 1000 K helium outlet earns the same 34.2 percent as a 530 K water outlet. That hides the entire argument for high-temperature reactors — the one thing an HTGR is for.

Shareable design URLs, so a student can send a teacher a link that opens their exact core. Currently, designs save to JSON files, which is one step too many for a classroom.

Built With

Share this project:

Updates

Submission history