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 2026, including ACTINV 1.1.2-era runs. They are not fresh measurements of every 1.3.1 feature or a ranking for every material, spectrum, and operating regime.

Three different comparison questions

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.

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.

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.

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.

Speed: numerical kernel versus a full calculation

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.

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 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.