Data and code: Total versus Updated Lagrangian forward models in machine-learning identification of concrete fracture parameters from notched-beam tests

This record holds the simulated databases, the digitized test curves, the identification results and the scripts used in the article “Total versus Updated Lagrangian forward models in machine-learning identification of concrete fracture parameters from notched-beam tests” (manuscript under review). Cite the article when you use the data or code. Contents data/simulated/doe_parameters.csv: The 400 Latin hypercube samples (seed 42) of E [Pa], f_t [Pa], f_c [Pa] and G_f [N/m]. data/simulated/dataset_TL_m96x12.csv, dataset_UL_m96x12.csv: 400 load–deflection curves per formulation (Total and Updated Lagrangian) on a 96 × 12 mesh with 48 load steps. The columns are the sample index, the four parameters, a complete flag (1 means the run reached the end; 398 of 400 per formulation), and P1–P30, the load [N] at the 30 deflections of the training grid (0.05–1.2 mm). data/simulated/slender_beam/: The study of when the formulations separate (Section 4.6). It covers elastic beams and CMOD-controlled notched beams with L/H = 10–80 in both formulations, a mesh study, and a step-size study of two outlier samples. data/experimental_digitized/: Load–deflection curves digitized from Petersson (1981), Report TVBM-1006, Lund Institute of Technology: Fig. 8.16 (two beams) and Fig. 7.14. DIGITIZING.md describes the method and the accuracy checks. The report pages themselves are not included. results/noise_study_metrics_96x12.csv: Identification metrics (RMSE, R², bias, coverage) by model, formulation and noise level (0, 1, 2 and 5% of peak load). It also covers whether models were tested on the formulation they were trained on or on the other one. results/phase8_identified.csv, phase8_cv_metrics.csv: Parameters identified from the three Petersson curves (GP and BNN, TL and UL) and the matching cross-validation metrics. results/phase8_forward_curves/: Forward re-analyses with the identified and the reference parameters on the 96 × 12 and 192 × 24 meshes. code/forward/: nlfea.py is the 2-D nonlinear FE solver (TL and UL, de Vree damage with crack-band G_f regularization). gen_database.py, run_all.ps1 and merge_shards.py generate the database in parallel and can resume after interruption. verify_03_energy.py is the energy check. code/inverse/: Gaussian process, neural network and deep-ensemble (BNN) inverse models, plus the cross-identification and noise study. code/validation_phase8/: Identification from the measured curves and the forward checks. code/slender_beam/: The slender-beam and CMOD-controlled solver (slender.py) and its run scripts. code/digitizing/: Pixel-tracing scripts for the Petersson figures. They need the page images (see DIGITIZING.md). code/matlab/: MATLAB TL/UL verification. TL and UL agree to 2e-12 for small strain, and the results match beam theory within 1.4%. Licenses: data CC BY 4.0, code MIT.