PWR-HYBRID-3/code/scripts/barrier/barrier_sos_2d.jl
Dane Sabo c5133401e0 Session work scratch: scram X_exit refactor, hot-standby SOS, fat scram tubes, model cheatsheet, journal entry
Multi-session work bundle on a draft branch.  Splits into a clean
sequence of commits later; pushed here so it isn't lost on a reboot.

Reach work
- code/scripts/reach/reach_scram_pj.jl: shutdown_margin halfspace
  X_exit (replaces "n <= 1e-4 AND T_f bound" framing); per-step
  envelope extraction added.
- code/scripts/reach/reach_scram_pj_fat.jl: per-step envelope
  extraction added; shutdown_margin discharge logic mirrored from the
  tight scram script.  3 probes (10/30/60s) all discharge from the
  fat union polytope.
- code/scripts/reach/reach_scram_full_fat.jl (NEW): full nonlinear
  PKE scram reach with fat entry.  Hits the stiffness wall at
  ~1.5 s plant time as expected; saves NaN-tolerant per-step
  envelopes.  Demonstrates concretely why PJ is the right tool for
  the longer-horizon proof.
- code/scripts/reach/reach_heatup_pj.jl: T_REF_START_C constant
  (entry-conditioned ramp) replaces T_STANDBY-init that was making
  the FL controller command cooling at t=0.  Per-step extraction
  already in place.
- code/configs/heatup/tight.toml: bumped maxsteps; probe horizon
  parameterized.

Hot-standby SOS barrier
- code/scripts/barrier/barrier_sos_2d_shutdown.jl (NEW): mirrors the
  operation SOS machinery on the hot-standby thermal projection.
  Includes the eps-slack pattern (so feasibility doesn't silently
  collapse to B == 0).
- code/scripts/barrier/barrier_sos_2d.jl: refactored to use the same
  helper.
- code/src/sos_barrier.jl (NEW): solve_sos_barrier_2d helper module
  factoring out the SOS construction; eps-slack with eps_cap=1.0 to
  avoid unbounded primal.

Library
- code/src/pke_states.jl (NEW): single source of truth for canonical
  initial-condition vectors per DRC mode (op, shutdown, heatup) keyed
  off plant + predicates.
- code/scripts/sim/{main_mode_sweep,validate_pj}.jl, code/CLAUDE.md:
  migrated to pke_states.

Predicates + invariants
- reachability/predicates.json: new shutdown_margin predicate (1%
  dk/k tech-spec floor, expressed as alpha_f*T_f + alpha_c*T_c
  halfspace).  Used as scram X_exit.

Plot script
- code/scripts/plot/plot_reach_tubes.jl: plot_tubes_scram_pj() with
  variant=:fat|:tight knob; plot_tubes_scram_full() for full-PKE
  3-panel (T_c, T_f, rho); plot_tubes_heatup_pj() reads results/
  not reachability/.

Journal + memory
- journal/entries/2026-04-27-shutdown-sos-and-scram-X_exit.tex (NEW):
  long-form entry on the SOS hot-standby barrier and the scram X_exit
  refactor.
- journal/journal.tex: input chain updated.
- claude_memory/ — three new session notes:
  * 2026-04-27-scram-X_exit-shutdown-margin.md
  * 2026-04-28-DICE-2026-conference-intel.md (people, sessions,
    strategic notes for the May 12 talk)
  * 2026-04-28-path1-sos-pj-sketch.md (sketch of nonlinear-SOS via
    polynomial multiply-through; saved for an overnight session)

Docs
- docs/model_cheatsheet.md (NEW): one-page reference of state vector,
  dynamics, constants, modes, predicates, sanity numbers — the talk
  prep cheatsheet Dane asked for.
- docs/figures/reach_*_tubes.png: regenerated with the new mat data.
- presentations/prelim-presentation/outline.md: revised arc per the
  April-28 review pass (cuts: Lyapunov-fails standalone slide,
  operation-tube standalone slide, SOS standalone; adds: scopes-of-
  control framing, scram on the headline result slide).
- app/predicate_explorer.jl: minor.

Hacker-Split: end-of-session scratch bundle
2026-05-02 23:02:50 -04:00

96 lines
3.5 KiB
Julia

#!/usr/bin/env julia
#
# barrier_sos_2d.jl — SOS polynomial barrier on a 2-state projection
# of the operation-mode LQR closed loop.
#
# Proof of concept that SumOfSquares.jl + CSDP can fit a polynomial
# barrier certificate on a reduced model. If this works, scaling to
# full 10-state is a matter of increasing degree and throughput.
#
# Reduced dynamics: project the LQR closed-loop onto (dn, dT_c), the
# dominant unregulated direction and the primary safety direction.
# A_red, B_red are the 2x2 / 2x1 submatrices corresponding to these
# components (ignoring cross-coupling into the 8 other states, which is
# a modeling simplification but keeps the SOS tractable).
#
# Safety: |dT_c| ≤ 5 K AND |dn| ≤ 0.15 (i.e. 0.85 ≤ n ≤ 1.15).
# Entry: |dT_c| ≤ 0.1 AND |dn| ≤ 0.01.
# Unsafe focus: dn ≥ +0.15 (high-flux trip; the harder direction
# under LQR because positive-n excursions trip n_high before T_c trips).
using Pkg
Pkg.activate(joinpath(@__DIR__, "..", ".."))
using LinearAlgebra
using MatrixEquations
using DynamicPolynomials
using SumOfSquares
using CSDP
include(joinpath(@__DIR__, "..", "..", "src", "pke_params.jl"))
include(joinpath(@__DIR__, "..", "..", "src", "pke_th_rhs.jl"))
include(joinpath(@__DIR__, "..", "..", "src", "pke_linearize.jl"))
include(joinpath(@__DIR__, "..", "..", "src", "sos_barrier.jl"))
plant = pke_params()
# Full linearization at full-power steady state.
A_full, B_full, B_w_full, _, _, _ = pke_linearize(plant)
# Reduced 2x2: rows/cols (1, 9) — n and T_c.
reduce_idx = [1, 9]
A_red = A_full[reduce_idx, reduce_idx]
B_red = B_full[reduce_idx]
B_w_red = B_w_full[reduce_idx]
# LQR on the reduced system. Light weighting on n, heavy on T_c.
Q_lqr = Diagonal([1.0, 1e2])
R_lqr = 1e6 * ones(1, 1)
X_ric, _, _ = arec(A_red, reshape(B_red, :, 1), R_lqr, Matrix(Q_lqr))
K_red = (R_lqr \ reshape(B_red, 1, :)) * X_ric
A_cl_red = A_red - reshape(B_red, :, 1) * K_red
# Cross-coupling check from dropped states.
cross = A_full[reduce_idx, setdiff(1:10, reduce_idx)]
println("\n=== SOS barrier — 2-state (dn, dT_c) projection of operation LQR ===")
println(" A_cl_red =")
show(stdout, "text/plain", A_cl_red); println()
println(" B_w_red = $B_w_red")
println(" eigenvalues: ", round.(eigvals(A_cl_red); sigdigits=4))
println(" ‖dropped-coupling‖ = $(round(norm(cross); sigdigits=3))")
println()
# --- SOS sets ---
@polyvar x1 x2 # x1 = dn, x2 = dT_c
entry_halfspaces = [
0.01 - x1, # dn ≤ 0.01
x1 + 0.01, # dn ≥ -0.01
0.1 - x2, # dT_c ≤ 0.1
x2 + 0.1, # dT_c ≥ -0.1
]
# Unsafe focus: dn ≥ +0.15 (high-flux trip). Asymmetric — n_high trips
# at 1.15 (dn = +0.15), n_low at 0.15 (dn = -0.85), so the +0.15
# direction is the binding one for LQR which has tightly bounded n.
unsafe_halfspaces = [x1 - 0.15]
# --- Solve ---
println(" Solving SOS feasibility (degree-4 B, ε-slack capped at 1.0)...")
result = solve_sos_barrier_2d(A_cl_red, (x1, x2),
entry_halfspaces, unsafe_halfspaces;
barrier_degree=4, multiplier_degree=2,
eps_cap=1.0)
println(" Status: $(result.status)")
if result.status == MOI.OPTIMAL && result.ε > 1e-8
println(" ✅ ε* = $(round(result.ε; digits=4)) — real certificate.")
println(" B(x) = $(result.B)")
elseif result.status == MOI.OPTIMAL
println(" ⚠ ε ≈ 0 — solver returned trivial B ≡ 0. No real barrier")
println(" at degree 4 with these sets.")
else
println("$(result.status). Try higher degree or relax sets.")
end