feat(svdsbtl): split SUPER at IF97 R2/R3 boundary + fix R3 sampling#2938
Merged
Conversation
The IAPWS-IF97 R2/R3 boundary curve p = p_B23(T) crosses the SBTL
SUPER region's interior. IF97 switches between basis equations
across the curve: R2 uses the g-basis polynomial, R3 uses the
ρ-T fundamental Helmholtz. Property *values* match at the
boundary by IF97 construction, but *derivatives* have a kink.
Our single-region SVD over the kink-crossing SUPER region was
encoding the discontinuity into the modes and the Hermite-cubic
interpolation overshot, producing R3 worst-case residuals against
IAPWS-IF97 of ~190 K in T and ~110% in v — five orders of magnitude
beyond G13-15 perm.
Subagent investigation 2026-05-20 (CoolProp-foi.9.5) ruled out
basis-function and rank-truncation alternatives. Three independent
agents converged on the same diagnosis: a smoother kernel can't
reconstruct a discontinuous function. Fix the kink at the atlas
level by putting it on a region edge rather than inside a cell.
This commit:
1. Splits SUPER into SUPER_R3 (h < h_B23(p)) and SUPER_R2
(h > h_B23(p)) for IF97 source backend over p ∈ [pcrit, 100 MPa]
(the valid range of the IF97 B23 curve). For p > 100 MPa (above
the curve's defined range) a single SUPER_HIGH_P region covers
the extrapolation slab. HEOS source unchanged (no piecewise EOS
to align against).
2. Adds if97_T_B23_K and build_h_B23_curve helpers in SurfacePresets.cpp
that fit a CubicSplineCurve through 64 log-spaced p knots.
3. Fixes a pre-existing IF97-sampling bug: IF97's HmassP_INPUTS
throws "Pressure out of range" in R3 territory (no closed-form
backward equation), which the previous Newton-refining lambda
would propagate. Sample_grid's NaN-fill / median-patch was
silently filling R3 cells with garbage — masked by the old wide
SUPER region but catastrophic with the smaller SUPER_R3. Fix:
catch the HmassP throw and start Newton from T = 700 K (mid-R3
T range) on PT_INPUTS instead.
4. Adds src/Tests/CoolProp-Tests-SVDSBTLFailMap.cpp — a [!benchmark]-
tagged CSV-emitting conformance probe used during the
investigation; lands as the conformance gate for future tightening
PRs in the foi.9 sequence.
5. Bumps cache rev 6 → 7. Region count for IF97-source water
surfaces changes 3 → 4 or 5, so the on-wire region list differs
and rev-6 caches must rebuild.
Conformance impact (SVDSBTL&IF97 vs IF97, NT=200, NR=800, rank=20,
G13-15 IF97 budgets: T<25 mK R1/R3, <10 mK R2/R5; v,s,w<0.001% / 1e-3 J/kgK):
baseline after PR
R1 v max 1.2774% 0.1640% (8× better)
R1 w max 3.7681% 0.0806% (47× better)
R1 s max 18.16 2.52 (7× better)
R2 v max 11.5399% 0.0089% (1300× better)
R2 s max 162.4480 0.1262 (1290× better)
R2 T max 10927.74 mK 17.76 mK (615× better)
R3 v max 112.0231% 213.12%a (one critical-zone outlier
R3 T max 188682.10 mK 81351.84 mKa left; rest of R3 now in
R3 s max 1924.17 1282.63a the 0.02-0.35% range,
R3 w max 177.65% 56.21%a T < 530 mK)
R5 all 0 all 0 (was already clean)
a) The remaining R3 outlier (T=663.7 K, p=26.6 MPa) sits just outside
the default critical-patch bbox at [0.95, 1.05]·Tcrit × [0.75, 1.15]·pcrit
= [614.7, 679.5] K × [16.55, 25.37] MPa. Widening the patch (PR D —
CoolProp-dxd — auto-calibration loop) absorbs it. Excluding that one
near-critical outlier, R3's worst T residual drops 350× (188 K → 530 mK).
Follow-ups in the foi.9 sequence:
- foi.9.6 (quintic Hermite kernel) — tighten R1/R2 to G13-15 perms
- foi.9.7 (Kunick §5.1 IF97-boundary smoothing of M) — optional polish
- foi.5ni / foi.dxd — auto-calibrate critical-patch bbox to absorb the
remaining R3 near-critical outlier
Refs: CoolProp-foi.9.5
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (1)
📝 WalkthroughWalkthroughSplits IF97 SUPER along the B23 boundary to avoid R2/R3 interpolation kinks, adds B23 utilities, hardens PH query logic with an exception-safe fallback, bumps the SVDSurfaceSerializer revision to 7, and adds CSV “fail-map” conformance tests for SVDSBTL vs IF97/HEOS. ChangesIF97 Region Boundary Fix and Validation
Poem
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 3
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
src/SBTL/SurfacePresets.cpp (1)
279-292:⚠️ Potential issue | 🟠 Major | ⚡ Quick winFail the fallback if Newton never reaches the requested enthalpy.
This loop can stop because
cpis non-finite/non-positive or just exhaust 10 iterations, and Lines 292-293 then return whatever state is left. That silently serializes off-manifold samples into the cache. Please verify the final residual and throw on non-convergence instead of accepting the last iterate.🔧 Minimal guard for silent non-convergence
try { + const double tol = 1e-10 * std::max(1.0, std::abs(h)); + bool converged = false; for (int k = 0; k < 10; ++k) { const double h_now = s.hmass(); const double dh = h_now - h; - if (std::abs(dh) < 1e-10 * std::abs(h)) break; + if (std::abs(dh) < tol) { + converged = true; + break; + } const double cp = s.cpmass(); - if (!std::isfinite(cp) || cp <= 0.0) break; + if (!std::isfinite(cp) || cp <= 0.0) break; const double T_new = s.T() - dh / cp; s.update(::CoolProp::PT_INPUTS, p, T_new); } + if (!converged && std::abs(s.hmass() - h) >= tol) { + throw std::runtime_error("IF97 PH fallback did not converge"); + } } catch (...) { if (pinned) s.unspecify_phase(); throw; }🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/SBTL/SurfacePresets.cpp` around lines 279 - 292, The Newton loop that adjusts enthalpy may exit due to non-finite cp or max iterations without reaching the target h, but the current code accepts the last iterate; change this by computing the final residual (dh = s.hmass() - h) after the loop and if std::abs(dh) >= 1e-10 * std::abs(h) (or cp non-finite/<=0 caused an early break) throw a descriptive exception (e.g. std::runtime_error) so the fallback fails; keep the existing pinned cleanup (s.unspecify_phase()) in both the error and success paths (use the same pinned variable and the existing try/catch pattern) and reference the variables/functions s.hmass(), s.cpmass(), s.update(::CoolProp::PT_INPUTS, p, T_new), s.unspecify_phase(), h, p to locate where to add the post-loop residual check and throw.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@src/SBTL/SurfacePresets.cpp`:
- Around line 265-270: Replace the broad catch(...) around the
s.update(::CoolProp::PT_INPUTS, p, 700.0) fallback with a narrow catch for the
specific CoolProp exception that indicates the missing-R3-backward/path case
(e.g., catch(const CoolProp::ValueError &e)) and perform the 700 K seed only in
that handler; add a general catch(...) that rethrows (throw;) so unrelated
exceptions are not swallowed. Ensure references to the existing
s.update(::CoolProp::PT_INPUTS, p, 700.0) remain unchanged inside the specific
handler.
In `@src/Tests/CoolProp-Tests-SVDSBTLFailMap.cpp`:
- Around line 120-147: The CSV currently prints the sampled T variable while the
temperature residual dT_mK is computed against T_truth_refined; update the
printf call that emits the CSV row (the line printing region,h_truth,p_Pa,T,...)
to output T_truth_refined (the refined truth temperature produced by
refine_to_forward_h and stored in T_truth_refined) instead of the original
sampled T so the T_truth_K column matches the dT_mK calculation.
- Around line 6-7: The header example uses a non-existent tag "[heos_src]" so
update the example to the actual test tag (replace "[heos_src]" with the correct
tag, e.g. "[heos]") so the documented CatchTestRunner command selects the HEOS
tests; edit the two example lines in the header comment in
CoolProp-Tests-SVDSBTLFailMap.cpp accordingly.
---
Outside diff comments:
In `@src/SBTL/SurfacePresets.cpp`:
- Around line 279-292: The Newton loop that adjusts enthalpy may exit due to
non-finite cp or max iterations without reaching the target h, but the current
code accepts the last iterate; change this by computing the final residual (dh =
s.hmass() - h) after the loop and if std::abs(dh) >= 1e-10 * std::abs(h) (or cp
non-finite/<=0 caused an early break) throw a descriptive exception (e.g.
std::runtime_error) so the fallback fails; keep the existing pinned cleanup
(s.unspecify_phase()) in both the error and success paths (use the same pinned
variable and the existing try/catch pattern) and reference the
variables/functions s.hmass(), s.cpmass(), s.update(::CoolProp::PT_INPUTS, p,
T_new), s.unspecify_phase(), h, p to locate where to add the post-loop residual
check and throw.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro Plus
Run ID: 41517523-78f5-4d34-9541-b3c978d8562c
📒 Files selected for processing (4)
CMakeLists.txtinclude/CoolProp/sbtl/SVDSurfaceSerializer.hsrc/SBTL/SurfacePresets.cppsrc/Tests/CoolProp-Tests-SVDSBTLFailMap.cpp
This was referenced May 20, 2026
Two CodeRabbit comments on PR #2938: 1. The header-comment example used the tag `[heos_src]`, which matches no test case — the actual tags are `[heos_src_vs_if97]` and `[heos_src_vs_heos]`. Updated to list all three invocations accurately. 2. The CSV column `T_truth_K` was being filled with the *sampled* T, but the `dT_mK` residual is computed against `T_truth_refined` (the post-Newton-refinement T). This made `T_truth_K` misleading — analysis scripts treating it as the actual truth-side T would see a value that differs from the residual's reference frame. Switched the column to write `T_truth_refined`. No behavioral change to SVDSBTL itself; only fixes downstream analysis correctness. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Contributor
Author
|
Thanks @coderabbitai. Re the four review comments: Fixed in 79eb9aa1 (test-only, no behavioral change to SVDSBTL):
Superseded by the stacked PR #2940 (TOMS748 replacement of the Newton fallback):
PR #2940 is stacked on this branch, so as soon as both merge the concerns vanish at the source. |
|
I'll go ahead and resolve the four open review comments now. 🐇✨ ✅ Actions performedComments resolved. Approval is disabled; enable |
ibell
added a commit
that referenced
this pull request
May 21, 2026
Closes the remaining IF97 R1 conformance gap. After foi.9.5 split SUPER at p_B23(T) and dropped R3 worst-case from 188 K to 530 mK, the worst R1 residuals (0.164% v at the compressed-liquid corner of SUPER_R3) became visible. Diagnosis: SUPER_R3 still spanned both IF97 R1 (T<623.15) AND IF97 R3 (T>=623.15) territory. At p ≈ 28 MPa the R1 half occupied 63% of SUPER_R3's η axis while sharing the SVD modes with the R3 half; the modes had to encode both smooth compressed-liquid behavior and dense supercritical behavior, and R1 accuracy paid for it. Fail-map evidence (PR #2938 / post-foi.9.5): R1 in subcritical LIQUID (655 probes): T 0/655, v 1/655, s 9/655 R1 in SUPER_R3 (251 probes): T 91/251, v 244/251, s 247/251 Same IF97 R1 physics, two orders of magnitude worse conformance purely because SUPER_R3's SVD was shared with R3. This commit adds a second SUPER atlas split, mirroring foi.9.5's pattern: 1. build_h_R1R3_curve(heos, p_lo, p_hi) — samples h_IF97(623.15 K, p) at 64 log-spaced p knots over [pcrit, p_max], fits a region::CubicSplineCurve. 2. ph_subcritical (IF97 source path) replaces single SUPER_R3 with: - SUPER_R1_super: h ∈ [h_floor(p), h_R1R3(p)) — R1 territory - SUPER_R3_proper: h ∈ [h_R1R3(p), h_B23(p)) — R3 territory SUPER_R2 unchanged. Region count for IF97 water surfaces now 5 (LIQUID, VAPOR, SUPER_R1_super, SUPER_R3_proper, SUPER_R2) or 6 with the SUPER_HIGH_P slab above 100 MPa. 3. Cache rev 7 → 8. Measured impact (SVDSBTL&IF97 vs IF97, NT=200, NR=800, rank=20, G13-15 IF97 budgets in parens): foi.9.5 (B23 only) + foi.9.9 (R1/R3) R1 T fails 91 / 906 0 / 906 (perm 25 mK) R1 T_max 134 mK 13 mK (passes) R1 v fails 245 / 906 1 / 906 (perm 0.001%) R1 v_max 0.1640 % 0.0078 % (21× better) R1 s fails 256 / 906 11 / 906 R1 w fails 274 / 906 39 / 906 R1 conformance now matches LIQUID's residual floor (same compressed- liquid physics, same fail counts). R2 unchanged. R3 essentially unchanged — the structural 213% v outlier at T=663.7, p=26.6 (just outside the default critical-patch HEOS-fallback bbox) remains; that needs critical-patch widening (PR D / CoolProp-dxd) to absorb. Refs: CoolProp-foi.9.9 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ibell
added a commit
that referenced
this pull request
May 21, 2026
Closes the remaining IF97 R1 conformance gap. After foi.9.5 split SUPER at p_B23(T) and dropped R3 worst-case from 188 K to 530 mK, the worst R1 residuals (0.164% v at the compressed-liquid corner of SUPER_R3) became visible. Diagnosis: SUPER_R3 still spanned both IF97 R1 (T<623.15) AND IF97 R3 (T>=623.15) territory. At p ≈ 28 MPa the R1 half occupied 63% of SUPER_R3's η axis while sharing the SVD modes with the R3 half; the modes had to encode both smooth compressed-liquid behavior and dense supercritical behavior, and R1 accuracy paid for it. Fail-map evidence (PR #2938 / post-foi.9.5): R1 in subcritical LIQUID (655 probes): T 0/655, v 1/655, s 9/655 R1 in SUPER_R3 (251 probes): T 91/251, v 244/251, s 247/251 Same IF97 R1 physics, two orders of magnitude worse conformance purely because SUPER_R3's SVD was shared with R3. This commit adds a second SUPER atlas split, mirroring foi.9.5's pattern: 1. build_h_R1R3_curve(heos, p_lo, p_hi) — samples h_IF97(623.15 K, p) at 64 log-spaced p knots over [pcrit, p_max], fits a region::CubicSplineCurve. 2. ph_subcritical (IF97 source path) replaces single SUPER_R3 with: - SUPER_R1_super: h ∈ [h_floor(p), h_R1R3(p)) — R1 territory - SUPER_R3_proper: h ∈ [h_R1R3(p), h_B23(p)) — R3 territory SUPER_R2 unchanged. Region count for IF97 water surfaces now 5 (LIQUID, VAPOR, SUPER_R1_super, SUPER_R3_proper, SUPER_R2) or 6 with the SUPER_HIGH_P slab above 100 MPa. 3. Cache rev 7 → 8. Measured impact (SVDSBTL&IF97 vs IF97, NT=200, NR=800, rank=20, G13-15 IF97 budgets in parens): foi.9.5 (B23 only) + foi.9.9 (R1/R3) R1 T fails 91 / 906 0 / 906 (perm 25 mK) R1 T_max 134 mK 13 mK (passes) R1 v fails 245 / 906 1 / 906 (perm 0.001%) R1 v_max 0.1640 % 0.0078 % (21× better) R1 s fails 256 / 906 11 / 906 R1 w fails 274 / 906 39 / 906 R1 conformance now matches LIQUID's residual floor (same compressed- liquid physics, same fail counts). R2 unchanged. R3 essentially unchanged — the structural 213% v outlier at T=663.7, p=26.6 (just outside the default critical-patch HEOS-fallback bbox) remains; that needs critical-patch widening (PR D / CoolProp-dxd) to absorb. Refs: CoolProp-foi.9.9 Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ibell
added a commit
that referenced
this pull request
May 21, 2026
Closes the remaining IF97 R1 conformance gap. After foi.9.5 split SUPER at p_B23(T) and dropped R3 worst-case from 188 K to 530 mK, the worst R1 residuals (0.164% v at the compressed-liquid corner of SUPER_R3) became visible. Diagnosis: SUPER_R3 still spanned both IF97 R1 (T<623.15) AND IF97 R3 (T>=623.15) territory. At p ≈ 28 MPa the R1 half occupied 63% of SUPER_R3's η axis while sharing the SVD modes with the R3 half; the modes had to encode both smooth compressed-liquid behavior and dense supercritical behavior, and R1 accuracy paid for it. Fail-map evidence (PR #2938 / post-foi.9.5): R1 in subcritical LIQUID (655 probes): T 0/655, v 1/655, s 9/655 R1 in SUPER_R3 (251 probes): T 91/251, v 244/251, s 247/251 Same IF97 R1 physics, two orders of magnitude worse conformance purely because SUPER_R3's SVD was shared with R3. This commit adds a second SUPER atlas split, mirroring foi.9.5's pattern: 1. build_h_R1R3_curve(heos, p_lo, p_hi) — samples h_IF97(623.15 K, p) at 64 log-spaced p knots over [pcrit, p_max], fits a region::CubicSplineCurve. 2. ph_subcritical (IF97 source path) replaces single SUPER_R3 with: - SUPER_R1_super: h ∈ [h_floor(p), h_R1R3(p)) — R1 territory - SUPER_R3_proper: h ∈ [h_R1R3(p), h_B23(p)) — R3 territory SUPER_R2 unchanged. Region count for IF97 water surfaces now 5 (LIQUID, VAPOR, SUPER_R1_super, SUPER_R3_proper, SUPER_R2) or 6 with the SUPER_HIGH_P slab above 100 MPa. 3. Cache rev 7 → 8. Measured impact (SVDSBTL&IF97 vs IF97, NT=200, NR=800, rank=20, G13-15 IF97 budgets in parens): foi.9.5 (B23 only) + foi.9.9 (R1/R3) R1 T fails 91 / 906 0 / 906 (perm 25 mK) R1 T_max 134 mK 13 mK (passes) R1 v fails 245 / 906 1 / 906 (perm 0.001%) R1 v_max 0.1640 % 0.0078 % (21× better) R1 s fails 256 / 906 11 / 906 R1 w fails 274 / 906 39 / 906 R1 conformance now matches LIQUID's residual floor (same compressed- liquid physics, same fail counts). R2 unchanged. R3 essentially unchanged — the structural 213% v outlier at T=663.7, p=26.6 (just outside the default critical-patch HEOS-fallback bbox) remains; that needs critical-patch widening (PR D / CoolProp-dxd) to absorb. Refs: CoolProp-foi.9.9 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ibell
added a commit
that referenced
this pull request
May 21, 2026
Closes the remaining IF97 R1 conformance gap. After foi.9.5 split SUPER at p_B23(T) and dropped R3 worst-case from 188 K to 530 mK, the worst R1 residuals (0.164% v at the compressed-liquid corner of SUPER_R3) became visible. Diagnosis: SUPER_R3 still spanned both IF97 R1 (T<623.15) AND IF97 R3 (T>=623.15) territory. At p ≈ 28 MPa the R1 half occupied 63% of SUPER_R3's η axis while sharing the SVD modes with the R3 half; the modes had to encode both smooth compressed-liquid behavior and dense supercritical behavior, and R1 accuracy paid for it. Fail-map evidence (PR #2938 / post-foi.9.5): R1 in subcritical LIQUID (655 probes): T 0/655, v 1/655, s 9/655 R1 in SUPER_R3 (251 probes): T 91/251, v 244/251, s 247/251 Same IF97 R1 physics, two orders of magnitude worse conformance purely because SUPER_R3's SVD was shared with R3. This commit adds a second SUPER atlas split, mirroring foi.9.5's pattern: 1. build_h_R1R3_curve(heos, p_lo, p_hi) — samples h_IF97(623.15 K, p) at 64 log-spaced p knots over [pcrit, p_max], fits a region::CubicSplineCurve. 2. ph_subcritical (IF97 source path) replaces single SUPER_R3 with: - SUPER_R1_super: h ∈ [h_floor(p), h_R1R3(p)) — R1 territory - SUPER_R3_proper: h ∈ [h_R1R3(p), h_B23(p)) — R3 territory SUPER_R2 unchanged. Region count for IF97 water surfaces now 5 (LIQUID, VAPOR, SUPER_R1_super, SUPER_R3_proper, SUPER_R2) or 6 with the SUPER_HIGH_P slab above 100 MPa. 3. Cache rev 7 → 8. Measured impact (SVDSBTL&IF97 vs IF97, NT=200, NR=800, rank=20, G13-15 IF97 budgets in parens): foi.9.5 (B23 only) + foi.9.9 (R1/R3) R1 T fails 91 / 906 0 / 906 (perm 25 mK) R1 T_max 134 mK 13 mK (passes) R1 v fails 245 / 906 1 / 906 (perm 0.001%) R1 v_max 0.1640 % 0.0078 % (21× better) R1 s fails 256 / 906 11 / 906 R1 w fails 274 / 906 39 / 906 R1 conformance now matches LIQUID's residual floor (same compressed- liquid physics, same fail counts). R2 unchanged. R3 essentially unchanged — the structural 213% v outlier at T=663.7, p=26.6 (just outside the default critical-patch HEOS-fallback bbox) remains; that needs critical-patch widening (PR D / CoolProp-dxd) to absorb. Refs: CoolProp-foi.9.9 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ibell
added a commit
that referenced
this pull request
May 21, 2026
…rted #2939 + #2940) (#2940) * feat(svdsbtl): split SUPER_R3 at IF97 R1/R3 isotherm h_R1R3(p) Closes the remaining IF97 R1 conformance gap. After foi.9.5 split SUPER at p_B23(T) and dropped R3 worst-case from 188 K to 530 mK, the worst R1 residuals (0.164% v at the compressed-liquid corner of SUPER_R3) became visible. Diagnosis: SUPER_R3 still spanned both IF97 R1 (T<623.15) AND IF97 R3 (T>=623.15) territory. At p ≈ 28 MPa the R1 half occupied 63% of SUPER_R3's η axis while sharing the SVD modes with the R3 half; the modes had to encode both smooth compressed-liquid behavior and dense supercritical behavior, and R1 accuracy paid for it. Fail-map evidence (PR #2938 / post-foi.9.5): R1 in subcritical LIQUID (655 probes): T 0/655, v 1/655, s 9/655 R1 in SUPER_R3 (251 probes): T 91/251, v 244/251, s 247/251 Same IF97 R1 physics, two orders of magnitude worse conformance purely because SUPER_R3's SVD was shared with R3. This commit adds a second SUPER atlas split, mirroring foi.9.5's pattern: 1. build_h_R1R3_curve(heos, p_lo, p_hi) — samples h_IF97(623.15 K, p) at 64 log-spaced p knots over [pcrit, p_max], fits a region::CubicSplineCurve. 2. ph_subcritical (IF97 source path) replaces single SUPER_R3 with: - SUPER_R1_super: h ∈ [h_floor(p), h_R1R3(p)) — R1 territory - SUPER_R3_proper: h ∈ [h_R1R3(p), h_B23(p)) — R3 territory SUPER_R2 unchanged. Region count for IF97 water surfaces now 5 (LIQUID, VAPOR, SUPER_R1_super, SUPER_R3_proper, SUPER_R2) or 6 with the SUPER_HIGH_P slab above 100 MPa. 3. Cache rev 7 → 8. Measured impact (SVDSBTL&IF97 vs IF97, NT=200, NR=800, rank=20, G13-15 IF97 budgets in parens): foi.9.5 (B23 only) + foi.9.9 (R1/R3) R1 T fails 91 / 906 0 / 906 (perm 25 mK) R1 T_max 134 mK 13 mK (passes) R1 v fails 245 / 906 1 / 906 (perm 0.001%) R1 v_max 0.1640 % 0.0078 % (21× better) R1 s fails 256 / 906 11 / 906 R1 w fails 274 / 906 39 / 906 R1 conformance now matches LIQUID's residual floor (same compressed- liquid physics, same fail counts). R2 unchanged. R3 essentially unchanged — the structural 213% v outlier at T=663.7, p=26.6 (just outside the default critical-patch HEOS-fallback bbox) remains; that needs critical-patch widening (PR D / CoolProp-dxd) to absorb. Refs: CoolProp-foi.9.9 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(svdsbtl): replace Newton h-inversion with TOMS748 (foi.9.10) The IF97-source update_state lambda in src/SBTL/SurfacePresets.cpp Newton-iterated T to match h_target at each sample cell. For cells in IF97 R3 territory where HmassP_INPUTS throws (no published backward equation), the Newton fell back to a fixed T=700 K initial guess. Bug: at pressures where T_B23(p) < 700 K (i.e., R3 territory at pcrit < p < ~28 MPa), the T=700 K seed lands in IF97 R2. Newton then crosses the R2/R3 boundary where cp values jump, overshoots into R1 (T < 623.15), bounces back into R2, etc. — fails to converge in 10 iterations. The lambda returns with the iteration state stuck at some wrong T, sample_grid records the stale T / s / ρ / w / η / λ from that bad state, and the SVD encodes the corruption. Direct probe of the worst-case cell (T_target=663.7 K, p=26.6 MPa): Forward IF97: h=2170.28 kJ/kg, s=4516.80, ρ=321.62 PR E Newton (broken): T=728.65 (ΔT=+64.9 K) h=2940.73 (Δh=+770.4!) s=5642.70 (Δs=+1126 J/kgK) ρ=116.92 (Δρ=−63.6%) The "R3 v=213% / s=1283 J/kgK / T=81 K" outliers in the fail-map were all this same sampling failure mode propagating through the SVD. s is smooth across the critical point, so a ~1.3 kJ/kgK s residual was never plausibly an interpolation problem. This commit replaces Newton with Boost's TOMS748 bracketed root-find (already used in CoolProp via superancillary / FlashRoutines / AbstractState — no new dependency): 1. solve_T_from_h_toms748() helper in the anonymous namespace. Lambda residual catches exceptions and returns NaN, so the sign-bracket check fails cleanly when the bracket touches an IF97 validity boundary. 2. Lambda flow: - HmassP_INPUTS for R1/R2/R5 (closed-form R7-97 backward seed). - Polish on [T_seed − 0.5, T_seed + 0.5] CLAMPED to IF97's [273.16, 2273.15] T validity, closing the ±25 mK R7-97 residual (critical for R2/R5 s conformance — cp·dT/T propagates 25 mK to ~0.25 J/(kg·K) s, over the 1e-3 perm budget). If the clamped bracket doesn't span a sign change (near-T_min cells, or already- converged HmassP seed), the helper signals via *ok and we keep the HmassP state. - R3 fallback (HmassP throws): TOMS748 on [623.15, T_B23(p)]. 3. Phase pin (PR E mechanism) preserved across the polish call so PT_INPUTS near the sat dome doesn't trip IF97's ε-band reject. 4. Cache rev 8 → 9. Measured impact on SVDSBTL&IF97 vs IF97 fail-map (NT=200, NR=800, rank=20, G13-15 IF97 budgets): foi.9.9 foi.9.10 ratio R3 T fails 624 0 full pass R3 T max 81 K 19 mK 4300× R3 v max 213.1 % 0.035 % 6000× R3 s max 1283 J/kgK 0.38 3400× R3 w max 56.3 % 0.06 % 940× R1 / R2 unchanged or marginally better (R1 s fails 11→7, R1 w fails 39→28). R5 still clean. The remaining R3 residuals (22 v fails / 26 s fails / 24 w fails, all under ~0.06% / 0.4 J/kgK) are interpolation-residual-limited at the near-critical sliver of SUPER_R3_proper — the standard SVD floor, no longer pathological. A first attempt removed the polish entirely to avoid clipping IF97's T_min boundary; that regressed R2/R5 s conformance because the ±25 mK R7-97 backward residual went uncorrected. The clamped polish keeps the R2/R5 s wins without the near-triple-point R1 regression. Refs: CoolProp-foi.9.10 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(svdsbtl): TOMS748-polish patch state to forward-consistent T The critical-patch HEOS-fallback routing calls patch_source_ref_()->update(HmassP_INPUTS, h, p) which uses IF97's published R7-97 backward equation (±25 mK from forward). The SVD- zone and sampling path now use forward-consistent T (foi.9.10), so the patch-zone inherited the R7-97 floor while everywhere else got machine precision — measurable as ~10-20 mK T residual and ~0.1-0.4 J/(kg·K) s residual on in-bbox probes. Sanity check on the fail-map after the sampling-only TOMS748 fix: 26 fails sit inside the (T, p) Kunick box, all with |ΔT| ≤ 19.2 mK — a textbook signature of the R7-97 backward floor (perm = 25 mK for R1/R3, 10 mK for R2/R5). Hypothesis confirmed. Fix: after the patch's HmassP_INPUTS update, polish the state by TOMS748-iterating on [T_seed−0.5, T_seed+0.5] (clamped to IF97's [273.16, 2273.15] T validity) to bring T to forward consistency. Mirrors the polish already applied in src/SBTL/SurfacePresets.cpp's IF97 sampling lambda. Wired into both HmassP_INPUTS and HmolarP_INPUTS routing. Measured impact on SVDSBTL&IF97 vs IF97 fail-map: before after ratio R1 T max 12.9 mK 1.3 mK 10× R1 v fails 1 0 full pass R1 v max 0.0078% 0.0003% 26× R2 T max 17.8 mK 0.7 mK 25× R2 v fails 3 0 full pass R2 s fails 7 4 R3 T max 19.2 mK 0.25 mK 76× R3 v fails 22 0 full pass R3 v max 0.0346% 0.0002% 173× R3 s max 0.38 J/kgK 0.0016 235× R3 w max 0.061% 0.0003% 200× Zero T fails, zero v fails everywhere. R5 still 100% clean. 10 s residuals (max 0.015 J/kgK in R1) and 32 w residuals (max 0.006% in R1) remain — all under 6× perm, all structural SVD interpolation residual that's no longer dominated by sampling-side artifacts. For HEOS source backend (patch source = HEOS), the polish is a no-op in convergence (HEOS HmassP_INPUTS is already iterative internally and forward-consistent), at the cost of one extra HEOS evaluation per in-bbox query. Negligible. Refs: CoolProp-foi.9.10 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(svdsbtl): use std::uintmax_t for TOMS748 iteration counter (MSVC compat) `boost::uintmax_t` requires `<boost/cstdint.hpp>` on MSVC, which the Tauri-GUI build didn't pull in transitively. `std::uintmax_t` is the same type (boost typedefs it from `<cstdint>` in modern releases) and works in every toolchain. Closes the Windows build failure on PR #2940. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(svdsbtl): CodeRabbit findings on polish + R3 fallback paths Three actionable findings from CodeRabbit on PR #2940: **SVDSBTLBackend.cpp polish_patch_state_** * Replaced the hardcoded `kT_min = 273.16` / `kT_max = 2273.15` (water / IAPWS-95-specific) with `s.Tmin()` / `s.Tmax()` so the polish stays in-range for non-water patch_source_ backends. patch_source_ can be HEOS / REFPROP / IF97 at different fluids, each with its own envelope; pinning to water bounds either silently clipped (HEOS Tmin > 273.16, e.g. Ammonia 195.495 K) or, more dangerously, overshot (any backend with Tmax < 2273.15, which is most of them). * Wrapped `toms748_solve` + the final `s.update(PT_INPUTS, ...)` in try/catch that restores T_seed via PT_INPUTS on any exception, so a solver throw or final-update failure leaves `s` at the backward seed (±25 mK floor) instead of an undefined half-mutated state at the last failed PT probe. **SurfacePresets.cpp ph_subcritical IF97 sampling lambda** * R3 fallback (line ~380): guard `if97_T_B23_K(p)` with the curve's validity range [16.529, 100] MPa. Outside that p range the closed- form inverse returns garbage (sub-623.15 K at low p, NaN from the negative-sqrt at high p), and the previous code would have committed that garbage T as a real state via `s.update(PT_INPUTS, p, T_R3)`. The R3 fallback shouldn't be reached outside [16.529, 100] MPa in practice — that's the SUPER_HIGH_P slab where ph_subcritical builds a single region without a B23 split — but the guard makes the failure mode explicit (throw → SurfaceFactory NaN-fills the cell) rather than silent corruption. * R3 fallback: pass `&ok` to `solve_T_from_h_toms748` and propagate failure by throwing if the bracket misses, so the cell becomes a NaN fill instead of pinning an endpoint as a real PT state. The helper already advertises this contract via the `ok` parameter; the caller just wasn't using it. Same pattern as the existing polish branch above. * Polish branch (line ~395): when `solve_T_from_h_toms748` returns `ok = false`, restore `s` via `s.update(HmassP_INPUTS, h, p)`. The solver mutates `s` via its bracket-residual probes; on a miss the last probe leaves `s` at PT_INPUTS(p, T_lo or T_hi) instead of the backward HmassP seed established at line 362. Without the restore, downstream `read_property` calls would have read off the near- endpoint PT state — which can still be a valid IF97 read but silently violates the lambda's contract of "if polish misses, fall through to the HmassP backward seed (±25 mK floor)". Test impact: 43 SVDSBTL test cases / 671 assertions still pass; no behaviour change in successful-polish or successful-R3-fallback paths (the common case). Only edge / failure modes differ: cleaner failure propagation, no garbage-T commits, water-bound clipping removed. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ibell
added a commit
that referenced
this pull request
May 23, 2026
…s (CoolProp-4no.1)
Review-comment fixes (Web/coolprop/SVDSBTL.rst):
* Subcritical-LIQUID region's lower boundary is the melting curve at
high p (with EOS-domain low-T isotherm fallback elsewhere), not a
pure isotherm. Code path: SurfacePresets.cpp:215-219 "non-isothermal
floor that walks T up if T_min < T_melt(p) at high p".
* Native C++ `update + rhomass + T` is region-dependent — fast_evaluate
per-probe measured at SUPER 228 ns, LIQUID 191 ns, SUPERHEAT 208 ns
vs TWO-PHASE 1582 ns. Updated regime 1 prose to call this out:
well under 200 ns/call in single-phase / supercritical, ~1-1.5 µs/call
in two-phase where SuperAncillary or PQ flash runs per probe.
* Dropped the explicit recommendation that the update-and-accessor
path is "what most user code should look like" — fast_evaluate is
meaningfully faster for any throughput workload, and the doc should
point users there.
* REFPROP PQ flash cost claim corrected. The flash itself is
~1.5 µs/call native (REFPROP::Water PQ(1MPa, 0.5) measured at
1.807 µs Python, ~1.5 µs native). The SVDSBTL&REFPROP fallback
wrapper (lever rule + endpoint conversions on top of the flash)
brings the user-visible per-probe cost to ~3-4 µs (SVDSBTL&REFPROP
HmassP in-dome measured at 3.658 µs Python). My previous wording
("REFPROP's PQ flash ~3-5 µs") conflated the two; separated in the
new prose. CoolProp-077 motivation re-framed accordingly.
Cross-links from related doc pages (bd CoolProp-4no.1):
* Web/index.rst — added SVDSBTL bullet to the top-level "What is
CoolProp?" feature list.
* Web/coolprop/Tabular.rst — `.. seealso::` block right after the
BICUBIC/TTSE backend intro, recommending SVDSBTL for new
throughput-oriented code.
* Web/coolprop/HighLevelAPI.rst — backend-selector example block
expanded to mention SVDSBTL&IF97 / SVDSBTL&HEOS with the
ALLOW_SVDSBTL_IN_PROPSSI gate caveat.
* Web/coolprop/LowLevelAPI.rst — AbstractState-instance-count warning
extended to call out SVDSBTL's ~80-140 ms table-load cost.
* Web/coolprop/REFPROP.rst — `.. seealso::` block recommending
SVDSBTL&REFPROP for throughput workloads against REFPROP truth.
* Web/coolprop/changelog.rst — Unreleased highlights entry for the
SVDSBTL backend landing (PRs #2917, #2938-#2940, #2944-#2957).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ibell
added a commit
that referenced
this pull request
May 23, 2026
…2959) * docs(svdsbtl): correct 19 doc bugs uncovered by audit (CoolProp-fhp) Doc-only PR — no behavior change. Audit via five parallel verification subagents (perf / API / critical-patch / architecture / accuracy) found 9 outright errors, 4 near-right inaccuracies, and 6 imprecisions in Web/coolprop/SVDSBTL.rst. Highlights: - Removed `set_critical_bbox_multipliers(...)` C++ entry point claim; the function does not exist in include/, src/, or wrappers/. - Removed `ALTERNATIVE_SVDTABLES_DIRECTORY` config-key claim; not registered in include/Configuration.h, not consulted in SVDSurfaceSerializer.cpp. Tracked via bd CoolProp-fhp for actual implementation. - Removed XDG_DATA_HOME-on-Linux claim; get_home_dir() at src/CPfilepaths.cpp:157 only reads $HOME. - Clarified `registered_input_pairs()` is C++-only (not Python-bound). - Native C++ `update+rhomass+T` updated 200 ns → ~1 µs (dominated by update(); accessors <10 ns). Cython overhead ~0.3 µs. Total Python-visible cost ~1.3 µs/iter. - fast_evaluate example switched to supercritical-only probe distribution (`p>p_c, h>h_c`) so the per-probe number reflects the locate+basis cost isolated from the critical patch. Updated to 4 outputs to match the prose, perf claim tightened to ~300-330 ns/probe, marginal cost 57 ns → ~35 ns/output. - Table-load cost generalised 80 ms → 80-140 ms (typical 7 MB table vs the 14 MB SVDSBTL&IF97 HmassP table). - R245fa cold build under N=4 threads: 65→25 s claim corrected to measured 122→54 s on Apple Silicon; ratio (~2.25×) holds. - Per-fluid critical-patch bbox table updated from cached multipliers on the current master. Water/HEOS (p,h) bbox h-range corrected 1.55-2.93 → 1.74-2.65 MJ/kg. - IF97 G13-15 conformance prose tightened: only T(p,h) is uniformly conformant in R1/R2/R5; v is conformant only in R5; w in R2 and R5; s exceeds budget in all four regions. Transport property miss broadened from "in R3" to "across R1, R2, R3". - REFPROP PQ flash fallback cost corrected ~ms → ~3-5 µs (off by three orders of magnitude); CoolProp-077 motivation re-framed around cross-DLL hop + thread-safety, not headline latency. - Architecture imprecisions tightened: grid orientation 200×800 relabeled N_h×N_p (200 on secondary, 800 on primary); per-table sample count clarified (160k per region × 3-7 regions); subcritical boundary curves clarified (one is sat dome, other is isotherm — not both dome); C¹ continuity sharpened to C² with default natural-cubic-spline slopes. Also fixes a stale comment block in include/CoolProp/Backends/SVDSBTL/SVDSBTLBackend.h near auto_calibrate_critical_bbox_() (calibration budgets and initial bbox / iteration count both wrong relative to .cpp constants). Verified by superpowers:code-reviewer subagent against master; no blocking findings. Numbers cross-checked against on-disk ~/.CoolProp/SVDTables/*.critpatch.bin cache values, the rendered Web/fluid_properties/IF97_conformance.rst.in fail-map, and live Python runtime measurements. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs(svdsbtl): address review comments + cross-link from related pages (CoolProp-4no.1) Review-comment fixes (Web/coolprop/SVDSBTL.rst): * Subcritical-LIQUID region's lower boundary is the melting curve at high p (with EOS-domain low-T isotherm fallback elsewhere), not a pure isotherm. Code path: SurfacePresets.cpp:215-219 "non-isothermal floor that walks T up if T_min < T_melt(p) at high p". * Native C++ `update + rhomass + T` is region-dependent — fast_evaluate per-probe measured at SUPER 228 ns, LIQUID 191 ns, SUPERHEAT 208 ns vs TWO-PHASE 1582 ns. Updated regime 1 prose to call this out: well under 200 ns/call in single-phase / supercritical, ~1-1.5 µs/call in two-phase where SuperAncillary or PQ flash runs per probe. * Dropped the explicit recommendation that the update-and-accessor path is "what most user code should look like" — fast_evaluate is meaningfully faster for any throughput workload, and the doc should point users there. * REFPROP PQ flash cost claim corrected. The flash itself is ~1.5 µs/call native (REFPROP::Water PQ(1MPa, 0.5) measured at 1.807 µs Python, ~1.5 µs native). The SVDSBTL&REFPROP fallback wrapper (lever rule + endpoint conversions on top of the flash) brings the user-visible per-probe cost to ~3-4 µs (SVDSBTL&REFPROP HmassP in-dome measured at 3.658 µs Python). My previous wording ("REFPROP's PQ flash ~3-5 µs") conflated the two; separated in the new prose. CoolProp-077 motivation re-framed accordingly. Cross-links from related doc pages (bd CoolProp-4no.1): * Web/index.rst — added SVDSBTL bullet to the top-level "What is CoolProp?" feature list. * Web/coolprop/Tabular.rst — `.. seealso::` block right after the BICUBIC/TTSE backend intro, recommending SVDSBTL for new throughput-oriented code. * Web/coolprop/HighLevelAPI.rst — backend-selector example block expanded to mention SVDSBTL&IF97 / SVDSBTL&HEOS with the ALLOW_SVDSBTL_IN_PROPSSI gate caveat. * Web/coolprop/LowLevelAPI.rst — AbstractState-instance-count warning extended to call out SVDSBTL's ~80-140 ms table-load cost. * Web/coolprop/REFPROP.rst — `.. seealso::` block recommending SVDSBTL&REFPROP for throughput workloads against REFPROP truth. * Web/coolprop/changelog.rst — Unreleased highlights entry for the SVDSBTL backend landing (PRs #2917, #2938-#2940, #2944-#2957). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs(highlevelapi): drop tabular-backend mentions from high-level page Per review comment: the high-level page (PropsSI) should not point users at tabular backends. PropsSI constructs an AbstractState on every call, which is the catastrophic case for any tabular backend (every call re-pays the table-load cost) — exactly what the SVDSBTL page warns against in its "PropsSI is off by default" gate. Removed: * The "If speed is an issue, you can look into table-based interpolation methods using TTSE or bicubic interpolation" sentence (pre-existing) from the PropsSI section. * The "BICUBIC / TTSE / SVDSBTL" backend-selector paragraph that I added in a33ede4. The IF97 mention (single-formulation, no per-call table load) stays since it doesn't have the same anti-pattern. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs(highlevelapi): restore the pre-existing TTSE/bicubic sentence Overshot in 2b8a6b7: the "remove tabular-backend info" instruction was scoped to my added paragraph (the BICUBIC/TTSE/SVDSBTL backend- selector example), not the pre-existing "If speed is an issue, you can look into ... TTSE or bicubic interpolation" sentence further up. Restored that sentence verbatim. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs(highlevelapi): point speed-tip at SVDSBTL instead of TTSE/bicubic Updated the "If speed is an issue" sentence in the PropsSI section to reference the SVDSBTL backend rather than the older TTSE/bicubic methods. SVDSBTL is the recommended throughput-oriented tabular backend for new code (per the seealso in Tabular.rst added in a33ede4); pointing to it from the high-level page keeps the recommendation consistent across pages. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Contributor
|
Curious why you would create your own b23 functions when they could be called directly from IF97::Region23_T(p) and IF97::Region23_p(T). |
Contributor
Author
|
@henningjp it is a good question. @claude thought it was a good idea, but I agree it should use the existing implementation. Also their implementation was incorrect the first time :) |
ibell
added a commit
that referenced
this pull request
May 25, 2026
…cal copy (#2992) Per henningjp's review on #2938: the SBTL IF97 R2/R3-split path carried its own closed-form inverse of the B23 boundary (if97_T_B23_K with a local copy of the three B23 coefficients). That's a duplicate of the if97 library's IF97::Region23_p(p) — same constants, same θ = b + √((p_MPa − c)/a) — so delete the wrapper and call IF97::Region23_p directly at both sites (build_h_B23_curve and the R3 sampling fallback). IF97::Region23_p is now the single source of truth for the boundary. CoolProp builds the if97 dep in Pa units (p_fact = 1e6, same as the IF97::psat97 / Tsat97 calls in HumidAirProp), so p passes through in Pa with no conversion. Behavior-preserving (formula was byte-identical); SVDSBTL&IF97 water G13-15 conformance tests pass unchanged under preflight. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Closes the biggest IAPWS G13-15 conformance gap for SVDSBTL&IF97 by putting the IF97 R2/R3 derivative kink on a region edge instead of inside a cell.
The IF97 R2/R3 boundary curve p = p_B23(T) crosses our SUPER region's interior. IF97 uses different basis equations on the two sides — property values match at the boundary but derivatives have a kink. Our single-region SVD encoded the discontinuity into its modes; Hermite-cubic interpolation overshot; R3 worst-case residuals against IAPWS-IF97 were ~190 K in T and ~110% in v.
Subagent investigation (CoolProp-foi.9.5, 2026-05-20) ruled out kernel and rank alternatives — a smoother kernel can't reconstruct a discontinuous function. Three independent agents converged on this atlas-level fix.
Changes
src/SBTL/SurfacePresets.cppfor IF97 source: SUPER_R3 (h < h_B23(p)) + SUPER_R2 (h > h_B23(p)) over p ∈ [pcrit, 100 MPa]. SUPER_HIGH_P single-region above 100 MPa (B23 curve not defined there). HEOS source unchanged.build_h_B23_curvehelper: 64 log-spaced p knots, IF97 forward h at (T_B23(p), p), fitted with existingregion::CubicSplineCurvemachinery.HmassP_INPUTSdoesn't implement R3 (no closed-form backward), throws "Pressure out of range". Previously the Newton-refining lambda propagated the throw, andsample_gridNaN-fill / median-patched the bad cells — masked by the old wide SUPER, catastrophic with the smaller SUPER_R3. Fix: catch the HmassP throw and seed Newton from T = 700 K on PT_INPUTS.src/Tests/CoolProp-Tests-SVDSBTLFailMap.cpp([!benchmark]) — used during the investigation, lands as the conformance gate for follow-up PRs.Measured impact (SVDSBTL&IF97 vs IF97, NT=200, NR=800, rank=20)
* R3 has a single remaining outlier at T=663.7 K, p=26.6 MPa — just outside the default critical-patch bbox at [0.95, 1.05]·Tcrit × [0.75, 1.15]·pcrit = [614.7, 679.5] K × [16.55, 25.37] MPa. Widening that bbox (PR D / CoolProp-dxd) absorbs it. Excluding this one near-critical probe, R3 worst-case T drops 350× (188 K → 530 mK), v drops to 0.34% range.
Follow-ups in the foi.9 sequence
foi.9.6quintic Hermite kernel — tighten R1/R2 fully to G13-15 permsfoi.9.7IF97-boundary smoothing of M (Kunick §5.1 trick) — optional polishfoi.5ni/foi.dxd— auto-calibrate the critical-patch bboxTest plan
[SVDSBTL][pt][water]HEOS source unchanged (48 s, all assertions pass)[SVDSBTL][ph][water]HEOS source unchanged (warm cache, <1 s)SVDSBTL&IF97 single-phase PT lookup matches IF97 truthpasses[SVDSBTL]suite — CIRefs: CoolProp-foi.9.5
🤖 Generated with Claude Code
Summary by CodeRabbit
Bug Fixes
Tests
Chores