sfr-tdd-mcp

A small, steady-state calculator for a single horizontal sodium heat pipe, exposed as an MCP tool (calculate_heatpipe_heat_transfer).

This is a portfolio demonstration. It is not a qualified design or safety code.

Background

My first attempt at this project (oklo-aurora-mcp) contained AI-invented sources and a fake benchmark. I rebuilt it here, tracing every constant to a scanned page in ANL/RE-95/2 and pulling test oracles directly from the printed tables. The breakdown of the original errors is in oklo-aurora-mcp/PROVENANCE.md.

Related MCP server: thermal-mcp-server

Scope

• Corridor: 625–750 °C saturation temperature, 50–750 W, horizontal (θ = 0°), steady state only. Inputs outside this corridor are rejected.

• Geometry (frozen):

  • Tube: Do = 19.05 mm, wall 1.0 mm

  • Wick: 1.0 mm annular, sintered stainless powder (d_p = 100 µm, ε = 0.65)

  • Vapor core: Dv = 15.05 mm

  • Lengths: Le = 1.0 m, La = 0.5 m, Lc = 1.0 m

• Outputs: mass flow, wall temperatures, ΔT, capillary margin, vapor Re and Mach number, and a flow regime computed from those numbers.

Physics and sources

Why other operating limits are excluded

Capillary pumping is the binding limit. Hand calculations at the worst case (625 °C), using this code's property functions:

Across the corridor, the vapor flow is laminar (Re_v < 800) and Ma < 0.09, so the incompressible laminar assumption holds.

Verification status (plain version)

• Properties: every ANL/RE-95/2 equation used by the solver (rho_l, rho_g, P_sat, dH_v, sigma, mu_l, k_l) was checked against the scanned report pages, and the citations give section, equation and page. Liquid and vapor density are also tested against printed Table 1.3-1 (p. 87) at 900 K and 1000 K.

• Layer 2: the tests check against an independent hand calculation at 500 W / 650 °C, done outside the codebase. This is a regression anchor, not a code-to-code benchmark against LANL HTPIPE or experimental data.

• No NQA-1 qualification is claimed.

Uncertainty

• Surface tension: ANL states +/-11% (2 sigma). The capillary margin scales with sigma, so the corridor minimum M_cap of 1.48 becomes about 1.32 at the low end. That is still above 1.

• Vapor viscosity: this is PROVISIONAL. ANL/RE-95/2 gives no vapor viscosity. A Chapman-Enskog estimate for monomer Na gives 1.56e-5 Pa s at 650 C, against the code's 2.06e-5. That is roughly a 25% spread. Sensitivity with mu_v varied from x0.70 to x1.15:

  • M_cap changes by less than 1%, because the vapor drop is about 2% of the total pressure drop. The capillary conclusion is insensitive to mu_v.

  • End-to-end delta_T at 500 W / 650 C ranges from 2.17 to 2.73 K, because R_vapor is about half of the thermal resistance. Treat delta_T as +/-15%.

How it was built

Built AI-paired. I set scope, architecture, and invariants; the model wrote tests first, then code. An independent model audit caught a mislabeled benchmark and an inconsistent wick spec, and both were fixed. See the git log, which keeps the red/green history.

Run

PYTHONPATH=src python3 -m pytest -q
PYTHONPATH=src python3 -m sfr_mcp.server