Data for: A physics law that helps for the wrong reason: bias cancellation in a crash-box surrogate model and how calibration fixes it

Description

Explicit finite element simulations (OpenRadioss) of axially crushed AA6061-T6 square crash boxes at wall thicknesses of 2-4 mm and impact velocities of 20-160 km/h, with the neural-surrogate results reported in the article. The deposit contains, for each of the 52 analyses (48 cases and four additional mesh-study analyses), the input deck, run logs and interface time history (fe_inputs.zip); the case-level results extracted from the simulations (p366_48_all.csv); the surrogate errors for the unconstrained model, the original, magnitude-corrected and calibrated crush laws and two placebo targets over 20 seeds, including a constraint-weight sweep (p389_2out_*.csv); and the scripts that train the surrogates and read out the results. README.md gives the reproduction commands and MANIFEST.csv the SHA-256 hash of every file.

Steps to reproduce

Unzip fe_inputs.zip to inspect or rerun the OpenRadioss analyses. To reproduce the surrogate results, place the scripts and p366_48_all.csv in one folder and run the commands in README.md (Python 3.10.11, NumPy 1.26.4, PyTorch 2.5.1; seeds 42-61 fixed). p396_cq_2out_campaign_readout_v1.py recomputes the tables and statistics from the *_raw.csv files.

Institutions

• Pusan National University

  Busan, Busan

Categories

Licence