Benchmarking data-driven material models on the classic Treloar dataset (source code and results)

This repository provides the code, databases, and generated results accompanying the paper “Benchmarking data-driven material models on the classic Treloar dataset.” The benchmark compares six material-model identification approaches on the Rubber20/Treloar dataset: EUCLID Material Fingerprinting (MF) Adaptive Material Fingerprinting (AMF) Physics-Augmented Neural Networks (PANN) Constitutive Artificial Neural Networks (CANN) Generalized-Invariant Constitutive Artificial Neural Networks (GI-CANN) The uploaded material supports the reproduction of the timing, initialization, hyperparameter-sensitivity, model-comparison, and energy/stress/Hessian benchmark studies reported in the accompanying publication. Files included CODEs.zip contains the complete Python source code for all six methods, the local Material Fingerprinting implementation, the Rubber20 data-processing and evaluation routines, and the central compare_all.py script used to execute the four benchmark studies. RESULTs.zip contains the results generated for the accompanying paper. These include raw and summarized timing data, neural-network initialization runs, EUCLID and AMF hyperparameter studies, prediction and accuracy comparisons, selected model parameters, energy/first Piola stress/Hessian benchmark data, figures, run configuration information, and issue logs. README.md provides installation instructions, a description of the individual benchmark studies, command-line options, and the structure of the generated results. DB_HEI.pkl is the hyperelastic incompressible material-fingerprint database required by the standard Material Fingerprinting method. DB_HEIIA.pkl is the hyperelastic incompressible isotropic adaptive material-fingerprint database required by Adaptive Material Fingerprinting. The two database files are method assets rather than precomputed Rubber20 results. They are read by the comparison code but are not modified by compare_all.py. Installation and execution Extract CODEs.zip, open a terminal in the extracted repository directory, and create a Python virtual environment: python -m venv .venv Activate the environment on Linux or macOS: source .venv/bin/activate On Windows: .venv\Scripts\activate Install the required dependencies: python -m pip install --upgrade pip python -m pip install -r requirements.txt On Windows, long-path support may need to be enabled before installing the dependencies. The code uses the Material Fingerprinting implementation included under: external/material_fingerprinting/ An unrelated PyPI package with the same import name should not be used. Before running the comparison, place the supplied database files at: external/material_fingerprinting/databases/DB_HEI.pkl external/material_fingerprinting/databases/DB_HEIIA.pkl Alternatively, the databases can be generated using the database-generation scripts included in the repository. From the repository root, run: cd external python make_database_HEI.py python make_database_HEIIA.py cd .. Database generation can require substantial runtime and memory. Verify that the resulting files are located in external/material_fingerprinting/databases/ before starting the comparison. Reproducing the archived results The results provided in RESULTs.zip were generated from scratch using: python compare_all.py \ --n-timing 100 \ --n-seeds 100 \ --euclid-lambdas 1e-5 \ --study4-num-samples 100 \ --study4-repetitions 100 The central script performs four stages: Runtime measurements for all six methods. Neural-network initialization studies for PANN, CANN, and GI-CANN. Hyperparameter studies for EUCLID and AMF, using uniaxial tension and equibiaxial tension data for identification and pure-shear data for validation. Evaluation of the selected models in terms of strain-energy density (W), first Piola stress (P), and Hessian (H). Each invocation recreates the comparison output under: RESULTS/rubber20_comparison/ No previously computed Rubber20 method results are loaded during execution. The Rubber20 dataset contains uniaxial-tension and equibiaxial-tension identification data and pure-shear validation data; it does not provide a separate independent test path.