[antoinepado]

Two fixes have been done.
First, in the abs_xsect variable 10E was used instead of 1E, this changes nothing in practice as this error cancelled out.
Secondly, the bigger change is the conversion from $$cm^2$$ to barns and from $$cm^2$$ to $$fm^2$$. Here is my understanding:
We have
$$1 barn = 10^{-28} m^2$$
and
$$1 femtometre = 10^{-15} m$$

To convert $$cm^2$$ to barns:
$$1 cm^2 = 10^{-2} m \cdot 10^{-2} m = 10^{-4} m^2$$
So we multiply by $$10^{24} \cdot 10^{-24}$$ both sides and we have
$$10^{24} \cdot 10^{-24} \cdot 1 cm^2 = 10^{24} \cdot 10^{-24} \cdot 10^{-4} m^2$$
$$1 cm^2 = 10^{24} \cdot 10^{-28} m^2$$
$$1 cm^2 = 10^{24} \cdot barn$$

And to convert $$cm^2$$ to $$fm^2$$:
$$1 cm^2 = 10^{-2} m \cdot 10^{-2} m = 10^{-4} m^2$$
So we multiply by $$10^{26} \cdot 10^{-26}$$ both sides and we have
$$10^{26} \cdot 10^{-26} \cdot 1 cm^2 = 10^{26} \cdot 10^{-26} \cdot 10^{-4} m^2$$
$$1 cm^2 = 10^{26} \cdot 10^{-30} m^2$$
$$1 cm^2 = 10^{26} \cdot fm^{2}$$

Please let me know if I have misunderstood something. Thank you.

[willend]

@antoinepado thanks once more, I think your corrections are right.

They do however shift output orders of magnitude so I will leave to @farhi to:

• validate
• judge when to change/merge

(Especially since we are right before a new release.)

[willend]

@antoinepado, @farhi says we should wait - putting your PR in draft to not merge by chance. ;-) Will make it to 3.10.1 or something like that.

[farhi]

Hi @willend and @antoinepado

I think we can wait for after the release. This PR will definitively change many test values by orders. It can already be seen from the X/Test_PowderN examples:

• %Example: index=1 Detector: Sph_mon_I=1.29869e-12 (PowderN)
• Example: index=2 Detector: Sph_mon_I=3.88313e-14 (SX)
• %Example: index=3 Detector: Sph_mon_I=6.29534e-12 (FluoPowder)

The index=2 is for the SX, and is typically a factor 100 too small. This test is currently off.

[willend]

Review + estimated new targets (generated with Claude Code, checked by @willend).

Verdict: the fix is correct. One correction to the description: the old error did not cancel out. Downstream, abs_xlen = σ/V0 with V0 in ų and lengths in m, so σ in fm² gives m⁻¹ directly and the barns branch needs ×100. With the old factors:

• barns=1 (.hkl from CIF, SSRL): σ came out in fm² but was treated as barns, so absorption was 100× too strong.
• barns=0 (.lau): σ came out in barns but was treated as fm², so absorption was 100× too weak.

Measured shift. "Stock" is a single 1e6-ray run (seed 1000). "New target" is the inverse-variance weighted mean of 3 runs at 1e6 rays (seeds 1001, 2002, 3003) and 1 run at 1e7 rays (seed 4004), all with the fixed component:

Instrument / example	Old target	Stock	New target (weighted)	Status
SSRL_bl_11_2_white_src Etohit=6900	4.24366e+07	4.52e7	1.65837e+10 ± 8.0e7	active, updated
SSRL_bl_11_2_not_white_src Etohit=6900	1.3e+10	1.34e10	4.17271e+11 ± 2.8e9	value updated, still disabled
Test_PowderN index=2 (SX)	3.88313e-14	3.92e-14	1.91726e-12 ± 4.9e-15	value updated, still disabled
Test_FluoPowder E0=15 index=2 (SX)	1.5129e-19	1.53e-19	1.99226e-19 ± 4.6e-23	value updated, still disabled
Test_Mono Mono=6 (no target)	—	fpi 8.1e-14	fpi 3.9e-14 (single run)	n/a: emon is 0 in both runs

These target updates are pushed to this PR as commit 1abedf7b2.

SX examples kept disabled (commit 07170c60e): I had re-enabled Test_PowderN index=2 and Test_FluoPowder index=2, assuming they were off only because of this bug. They were not: they were disabled on purpose earlier, Test_PowderN index=2 in 3c9ec204b (Single_crystal powder mode numerically unstable on GPU) and Test_FluoPowder index=2 in cb9df2a79 (fails consistently on Windows, possibly cif2hkl UB). Windows failed both again on this PR. They are back to disabled, with the new values kept and a note in each file saying why.

%Scan references: the 31-point EnergyMonitor_first_I arrays in both SSRL instruments were also computed with the old absorption and failed CI (worst points off by 3227% and 35772%). Commit 91360680d regenerates them. Each point is the inverse-variance weighted mean of two 1e7-ray scans (seeds 11, 22); the seeds agree within 6% at every point. New ranges: white_src 2.0e9 to 6.8e9 (was ~1e7), not_white_src 3.2e11 to 4.9e11 (was ~1.4e10).

Unchanged: Test_PowderN index 1 and 3, Test_FluoPowder index 1, 3, 4 and 5, and every instrument using material_datafile="NULL" (Test_SX, Test_SX_standalone, SOLEIL_PX1, SOLEIL_PX2a).

Locally, this also fixes the "factor ~100 too small" SX result in Test_PowderN (index=2) that @farhi pointed out: it is now in line with the PowderN result (index=1), about 1.3e-12.

Open question (not verified, separate from this PR): abs_xsect is per atom (A[r] from Si.txt) but is divided by the volume of the whole unit cell (V0). For Si, with 8 atoms per cell, absorption may still be about 8× too weak.

Note: in the SSRL white_src runs the 1e6 results spread a little more than their error bars suggest (1.54e10 to 1.68e10), so CI's tolerance may need to absorb a few percent there.

[antoinepado]

@willend

Verdict: the fix is correct. One correction to the description: the old error did not cancel out.

I think the above sentence is wrong. We can see that the the error where 10E is used instead of 1E cancels out no?
Line 1546 a division by 10E23 was done, then a multiplication at lines 1552 and 1554 by either 10E26 or 10E24 is done.
This is the same as doing a division by 1E23, then a multiplication by either 1E26 or 1E24.
The error cancels out. Am I misunderstanding something?

The main fix is the corrected conversion though, not the 10E to 1E mistake that had no change in practice.