Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Understand the competitive benchmark

The competitive benchmark compares ACTINV with other inventory and depletion codes on accuracy, speed, and capability. It helps you judge which measured differences matter for an activation workflow. The full benchmark report retains the dated results and links to their evidence.

The comparisons below are recorded experiments from September to November 2026, including ACTINV 1.1.2-era runs, the ACTINV 1.7.0 shutdown-dose timing and the ACTINV 1.7.1 per-problem timing. They are not fresh measurements of every 1.7.1 feature or a ranking for every material, spectrum, and operating regime.

Three different comparison questions

ComparisonQuestion it answersWhat to check
Calculated responses against measurementsHow well does this code and selected data predict the experiment?Material, spectrum, history, response, and data versions
Codes using common evaluated cross sectionsHow much do processing, chain representation, and numerical methods differ?Decay data and represented product states can still differ
The same matrix/time step in numerical kernelsHow quickly do the implementations solve this numerical workload?Matrix size, call overhead, threading, hardware, and omitted setup work

Using TENDL-2025 in one code and TENDL-2017 in another compares both software and data choices. A common-cross-section comparison removes an important source of difference, but it does not make the two complete input chains identical.

Read the accuracy metrics

C/E is the calculated response divided by the experimental response. A value of 1 is exact agreement; 0.8 is 20% below the measurement and 1.2 is 20% above it.

MetricInterpretation
Median abs(ln(C/E))Typical multiplicative error across scored points; lower is better
90th percentile, or p90, abs(ln(C/E))Error toward the difficult end of the distribution; lower is better
Pooled geometric-mean C/EOverall multiplicative bias; closer to 1 is better
Experiments with all points inside the acceptance bandEntire measured cooling curves meeting the stated threshold; higher is better

The logarithm is natural. It treats a factor-of-two overprediction and a factor-of-two underprediction equally. An absolute log error of about 0.693 corresponds to a factor of two; it is not a 69.3% relative error.

The recorded FNS checker defines its “within 30%” band as abs(ln(C/E)) <= ln(1.3), or approximately 0.769 <= C/E <= 1.3. This is a multiplicative band, not the ordinary 0.7–1.3 arithmetic interval. An experiment counts only when every scored point meets it.

Read the metrics together. A geometric mean near 1 can hide large overpredictions and underpredictions that cancel in the average. Pooled point metrics give more weight to experiments with more scored time points; experiment-level metrics answer a different question. The checker records nonpositive or unaligned exclusions rather than putting them inside a logarithm.

FNS accuracy: ACTINV and FISPACT-II

The FNS benchmark contains 132 experiments across 73 materials. In the recorded common-TENDL-2017 comparison, 2,360 positive aligned point pairs were scored. ACTINV used ENDF/B-VIII.0 primary decay data with JEFF-3.3 fallback; the published FISPACT-II 4.0 reference used its own condensed decay dataset.

MetricACTINV / TENDL-2017FISPACT-II / TENDL-2017
Median absolute log error0.10300.1053
p90 absolute log error0.68940.6846
Pooled geometric-mean C/E1.06051.0636
Experiments with all scored points in the band71 / 13269 / 132

ACTINV has a small advantage on the typical error and whole-curve count in this arm; FISPACT has a slightly lower tail error. Switching ACTINV’s primary decay data to JEFF-3.3 changes its median error to 0.1058 and its passing count to 69. The narrow margin therefore depends on the decay-data choice; it does not establish a broad solver advantage independent of evaluation uncertainty.

Evidence: common-TENDL-2017 result and JEFF-primary sensitivity result.

The recorded ACTINV/TENDL-2025 comparison has median error 0.1392 and 59 passing experiments, while the same FISPACT/TENDL-2017 reference has 0.1053 and 69. That is a software-plus-data comparison. It shows why changing evaluations can matter more than a small numerical-method difference. Later decay-aware TENDL-2023 rebuilding improved a separate arm; its metrics and data identities are in the full report.

OpenMC: a common ENDF/B-VIII.1 comparison

This arm compares the activation/depletion responses on common evaluated ENDF/B-VIII.1 cross sections over 21 experiments and 424 measured points. ACTINV and OpenMC 0.15.3 use different chain representations and processing paths.

MetricACTINV / ENDF-8OpenMC / ENDF-8
Median absolute log error0.1310.181
p90 absolute log error0.9701.640
Pooled geometric-mean C/E0.7880.685
Experiments with all scored points in the band10 / 2110 / 21

ACTINV has smaller median and tail errors and bias closer to 1 in this subset. The whole-curve passing count is tied. The detailed comparison attributes important differences to represented isomer-production channels, including tantalum, tungsten, and yttrium. Those findings describe this activation setup; they do not compare transport accuracy or every OpenMC depletion model.

See the ENDF-8 comparison for per-experiment results and shared data-driven misses.

All 132 FNS experiments

Protocol P172 repeated the comparison on every FNS experiment, with OpenMC 0.15.3 in its published configuration and ACTINV both with its shipped data and with OpenMC’s cross sections. P179 reran ACTINV’s identical-data arm after the 1.7.1 capture repair.

Against 2360 measured decay-heat points from 132 JAEA FNS experiments, ACTINV with its shipped TENDL-2025 data has a median |ln C/E| of 0.139; OpenMC 0.15.3 depletion with its published ENDF/B-VIII.1 fast-spectrum depletion chain (chain_endfb81_fast.xml) and ENDF/B-VIII.1 has 0.316.

With its R-matrix-limited capture repaired and otherwise the same ENDF/B-VIII.1 cross sections as OpenMC, ACTINV has a median |ln C/E| of 0.234 on 2360 measured decay-heat points from 132 JAEA FNS experiments.

With isomer branching ratios from the ENDF/B-VIII.1 library added to that chain, OpenMC has a median |ln C/E| of 0.252 on the 2360 of those points it covers.

With a chain built from ENDF/B-VIII.0 decay data and TENDL-2025 reaction channels instead (Chain.from_endf, ground-state targets), OpenMC has a median |ln C/E| of 0.311 on the 2358 of those points it covers.

Evidence: P172 scored points and verdict and P179 verdict.

Activation reaction rates: a second facility

The FNS experiments above all come from one facility. Protocol P167 adds an independent one with open data: the 7Li(p,n) quasi-monoenergetic neutron source of the Nuclear Physics Institute of the Czech Academy of Sciences at Řež, where foils were irradiated in fields from 12.4 MeV and 14.4 MeV protons and the spectrum at the foils was measured (Matěj, Košťál, Majerle et al., arXiv:2306.09353). ACTINV computed each reaction rate with its shipped TENDL-2025 library and its own spectrum rebin and collapse (actinv inspect xs).

Against 29 measured reaction rates in two quasi-monoenergetic 7Li(p,n) neutron fields at NPI Řež (12.4 and 14.4 MeV protons; Matěj et al., 2023), ACTINV with its TENDL-2025 library has a median |ln C/E| of 0.038 and a geometric-mean C/E of 1.023.

On the same points the authors’ calculations give a median |ln C/E| of 0.047 with ENDF/B-VIII.0, 0.051 with JEFF-3.3 and 0.057 with JENDL-4, and 0.030 with IRDFF-II on the 27 points it covers.

25 of the 29 ACTINV values are within 2σ of the measurement.

FieldPointsACTINV median abs log errorACTINV within 2σ
12.4 MeV protons130.0399
14.4 MeV protons160.03516

IRDFF-II is a dosimetry library fitted partly to integral measurements like these, so it is a reference rather than a peer. The largest ACTINV deviation is ⁶⁰Ni(n,p) in the 12.4 MeV field, 18% high, where every library is high by about the same amount. The authors normalized the spectra to ⁵⁸Ni(n,p) and ²⁷Al(n,α), so those two reactions are not counted; ACTINV reproduces them within 1%. This checks activation cross sections between about 1 and 13 MeV; it does not check inventories or decay heat.

Evidence: reaction rates and verdict.

Fission decay heat: five laboratories

Fission-product decay heat tests the fission yields and decay data behind every fission inventory. Protocol P166 scored ACTINV against the measured tables of the IAEA CoNDERC fission compilation (Gauld, 2019): Friesenhahn 1976 and 1979 (IRT), Schrock 1978 (UC Berkeley) and Fiche 1976 (CEA Cadarache), for U-233, U-235 and Pu-239 thermal fission, with irradiations from a 1 s pulse to 35 days. ACTINV used ENDF/B-VIII.0 yields and decay data. The compiler’s ORIGEN calculations, printed in the same tables, used ENDF/B-VII.0 yields and ENDF/B-VII.1 decay data.

Against 692 measured total decay-heat points from 12 fission decay-heat experiments at IRT, UC Berkeley and CEA Cadarache (U-233, U-235 and Pu-239 thermal fission), ACTINV with ENDF/B-VIII.0 yields and decay data has a median |ln C/E| of 0.046 and a geometric-mean C/E of 0.960.

On the same points, ORIGEN with ENDF/B-VII.0 yields and ENDF/B-VII.1 decay data (Gauld, 2019) has a median |ln C/E| of 0.042 and a geometric-mean C/E of 0.962.

LaboratoryPointsACTINV median abs log errorORIGEN median abs log error
IRT (Friesenhahn)5300.0420.040
CEA Cadarache (Fiche)420.0660.065
UC Berkeley (Schrock)1200.0910.083

The two codes track each other closely, so the remaining differences come from the data and the measurements rather than the solver. The UC Berkeley scatter is mostly at cooling times the compiler flags as unreliable; limited to 600–10,000 s, the pooled ACTINV figure is 0.040. ACTINV’s separately measured gamma and beta heat (Friesenhahn) have median |ln C/E| of 0.042 and 0.057. The earlier P9 comparison covers the Oak Ridge (Dickens) and Los Alamos (Yarnell) U-235 measurements, with geometric-mean C/E 1.007 and 0.985.

Evidence: scored points and verdict.

Speed: numerical kernel versus a full calculation

A full shutdown-dose calculation with OpenMC

Protocol P163 timed OpenMC’s shutdown-dose (R2S) workflow on the steel-cube problem from model to photon source, with ACTINV doing the activation step against OpenMC 0.15.3 doing it in flux mode, its fastest documented setting. Both codes used 2 threads, the same seeds, and the same transport settings; OpenMC used ENDF/B-VIII.1, ACTINV TENDL-2025.

Mesh cellsACTINVOpenMC flux modeWorst-case ratio
810.7–10.8 s99.5–102.0 s9.2
6416.4–16.8 s138.4–139.2 s8.2
51264.5–66.4 s435.1 s6.5

ACTINV is at least 6.5x faster than OpenMC 0.15.3 flux mode at every measured size (8, 64, 512 mesh cells).

Against OpenMC 0.15.3 in its default tally mode (per-nuclide reaction-rate tallies, one energy group), one run at 8 mesh cells, ACTINV’s steps 1 and 2 are at least 42x faster; this is not OpenMC’s fastest mode.

The shutdown dose rates agreed within 2.1% at every size. The ratio falls as the mesh grows because ACTINV’s per-cell activation cost comes to outweigh OpenMC’s fixed start-up cost. Evidence: timing record and verdict. See Use ACTINV with OpenMC to run it.

Larger meshes with several materials

Protocol P181 and its Amendment A timed the same workflow on a model where mesh elements cross material boundaries: a tungsten sphere in CuCrZr, water and steel around a 14.1 MeV source, meshed into 8 to 32,768 elements. OpenMC splits each element by material, so the activation regions outnumber the elements. Each run had its own 10 GiB memory limit and 4-hour limit, two CPU threads, and the same transport; one run per size, on an Intel Core i3-N305.

Elements (activation regions)ACTINV, model to photon sourceOpenMC 0.15.3 flux modePeak working memory, ACTINV / OpenMC
8 (32)11.0 s112 s1.2 / 6.4 GiB
64 (136)13.4 s180 s1.2 / 6.4 GiB
512 (776)31.8 s604 s1.3 / 6.5 GiB
4,096 (5,120)155 s3,432 s1.4 / 7.2 GiB

Inside the same 10 GiB memory limit and 4-hour limit per run, ACTINV (with an ENDF/B-VIII.1 activation library) completed meshes up to 32,768 elements (36,208 regions) and OpenMC 0.15.3 in flux mode up to 4,096 elements (5,120 regions); at 13,824 elements OpenMC was stopped by its memory limit.

ACTINV used an ENDF/B-VIII.1 activation library, the evaluation OpenMC used, and its shutdown dose rate agreed with OpenMC’s within 4.4% at every size. The first run used ACTINV’s shipped TENDL-2025 data: its doses were 11–21% below OpenMC’s, a difference between the two evaluations for this steel and tungsten that the protocol’s 10% sanity bound does not allow, and that run is not quoted. The amendment that reran ACTINV on ENDF/B-VIII.1 was written after that result was known, and its ACTINV runs came a few hours after OpenMC’s. With TENDL-2025, ACTINV also completed every size but was slower at large meshes (949 s at 4,096 elements). Evidence: amendment record, verdict and P181’s record.

Per-problem time on the FNS problems

Protocol P180 timed the activation calculation alone, with no transport, on the 132 FNS problems of P172. Each problem ran in its own process, so every time includes start-up and data loading, on a quiet machine (Intel Core i3-N305) limited to two cores.

On the 132 JAEA FNS decay-heat problems of P172, each solved in its own process on the same two CPU cores with the same ENDF/B-VIII.1 cross sections, ACTINV 1.7.1 took a median of 0.61 s per problem from start to finish and OpenMC 0.15.3 depletion 18.1 s; the median per-problem ratio was 29.8.

Counting only OpenMC’s depletion step, without its start-up and cross-section collapse, the median per-problem ratio was 9.

With its shipped TENDL-2025 data, ACTINV took a median of 0.88 s per problem.

Evidence: timings and verdict.

The numerical kernel

The recorded CRAM-48 benchmark uses identical operators at the same Python-call boundary with both implementations limited to one thread. Its ratio is OpenMC median time / ACTINV median time: above 1 favors ACTINV and below 1 favors OpenMC.

Operator statesRecorded time ratio
3213.8
2562.64
1,0241.28
2,0480.98
4,0961.83

The advantage varies with workload size, with near parity at 2,048 states. Small operators make Python and call overhead a larger fraction of the measured time. These figures do not include the complete cost of installation, library processing, cache preparation, or every response calculation. They cannot be applied as a universal speedup to a desktop run or an uncertainty study.

Evidence: kernel timings and numerical agreement. Use your own complete workflow and comparable hardware when estimating turnaround time.

Capability and error-bar comparisons

Capability is a separate axis: product-state handling, uncertainty channels, source exports, self-shielding, reverse calculation, and design search may matter even when nominal heat predictions are similar. Check the current workflow guide and qualification scope for ACTINV’s implemented boundaries. A feature being available does not establish its accuracy for every application, and an unavailable competitor was not measured. The 2026-09-30 capability scorecard refresh updates the ACTINV column for gas production, the transport-tally flux uncertainty channel, lean mesh output, and impurity budgets; the competitor columns there remain the dated CB1 survey.

The report also describes discrepancy calibration on 70 FNS experiments with 62 experiments held out. The recorded two-standard-deviation interval covers 93.1% of holdout points. That is useful evidence about this calibrated corpus, not a guarantee of 95% coverage on a new material or spectrum. See calibration evidence and the uncertainty reference.

Apply the results to your study

Find the comparison closest to your material, reactions, spectrum, cooling times, and response. Check per-material outliers and represented channels, not just the pooled score. Keep software performance, evaluation accuracy, and feature coverage distinct when making a tool choice.

The benchmark supports the recorded comparisons. Your calculation still needs suitable data and an applicability assessment; see Validation evidence, Known data limitations, and Scope and qualification.