Benchmarks & Evidence

See the test, the comparison and the limits behind a result.

For engineering teams evaluating simulation and AI tools. A numerical check compares a calculation with a known answer; an experiment checks it against the physical world. Passing one does not automatically establish the other.

Phase-change heat storage: a documented example

Existing September 2026 results for an assumed 40 × 30 × 10 mm aluminium-silicon sample. The model follows heat spreading, melting and freezing, without liquid flow or a container.

Recorded numerical checks; not measured material performance
TestComparisonRecorded resultMeaning
Analytical verification31 simplified cases against exact mathematics434 / 434 checks passedChecks the numerical method with an invented test material.
Working-grid check33 versus 65 depth cells1.4207 K centre-temperature differenceWithin the declared 3 K limit. One kelvin of difference equals one degree Celsius.
Time-step checkWorking time step versus half that step0.0092 K centre-temperature differenceChecks sensitivity to how far each calculation advances in time.
Rejected coarse grid17 versus 33 depth cells5.7487 K centre-temperature differenceAbove the 3 K limit. The raw campaign remains FAIL: 219 of 220 checks passed.
Original grid and time-step comparison plots: the coarse 17-cell grid is rejected while the selected 33-cell grid passes its finer-grid check.
Original saved result figure. Open it to inspect the axes and labels at full resolution.

Hardware and repeatability

Recorded on NVIDIA Tesla V100 graphics processors. The saved material study gives identical histories and fields when the calculation is shared by one, two and four processors. This is a consistency check, not a claim of universal speedup.

What remains unproven

Independent tests on the actual alloy, complete storage-device efficiency and real equipment charging time are not established. Separate communication-software memory checks were not clean; passing numerical checks does not erase those limitations.

Full study and model boundariesRead the technical report (PDF)Download the scenario table (CSV)

Reading AI performance claims

Prediction error

How far the prediction is from its reference. Mean absolute error is the average size of a difference; mean squared error (MSE) squares differences before averaging. A percentage needs a stated normalisation and test set.

Error reduction

How much error falls between two model versions. A 98.94% reduction from a baseline is not 98.94% physical accuracy. Internal model labels alone do not define a reproducible test.

Speed

A runtime ratio needs the same task, hardware, problem size and timing boundary. Prediction after training is different from total training, data preparation and simulation cost.

Headline accuracy and speed figures elsewhere describe selected company-reported cases, not guarantees for every model. We do not combine them into one score. Request the matching test definition and evidence before applying a figure to your engineering decision.

Discuss a benchmark for your problemExplore AI demonstrations