Machine-learning-enhanced Arrhenius constitutive modeling, microstructural evolution and deformation mechanisms of Gd during hot compression
Description
This dataset contains the source data, training code and trained model files for the machine-learning flow-stress models developed for high-purity gadolinium (Gd, 99.86 wt.%) under hot compression. Flow-stress curves were obtained by isothermal uniaxial compression on a Gleeble-3500 thermomechanical simulator at 500-800 °C and strain rates of 0.01-10 s^-1, to a height reduction of 60%. All curves were corrected for friction and deformation heating before modelling. Three models are included: (1) BP-E, a direct-stress back-propagation neural network ensemble using true strain, a temperature descriptor and log10(strain rate) as inputs; (2) an Arrhenius-residual BP ensemble trained on the residual between the experimental stress and a fixed seventh-order strain-compensated Arrhenius baseline; (3) AR-FHE, a functional heterogeneous ensemble combining functional Gaussian process regression, functional gradient boosting regression and the Arrhenius-residual BP model, in which residual curves are represented by cubic B-splines and reduced by principal component analysis, with ensemble weights estimated from out-of-fold predictions under leave-one-curve-out cross-validation. Contents: training scripts, requirements files, reproducibility instructions, the 25 training curves, intermediate Arrhenius baseline tables, trained model weights, and prediction outputs. External-test isolation: the experimental curves at 650 °C and strain rates of 0.01, 0.1, 1 and 10 s^-1 were excluded from model training, hyperparameter optimization and ensemble-weight estimation. They are provided only as ground truth for the final external evaluation.
Files
Steps to reproduce
Environment ----------- Python 3.10 or later, CPU only. Install dependencies separately in each model directory: python -m pip install -r requirements.txt Run order --------- The AR-FHE ensemble consumes outputs of the residual-BP and BP-E models, so it must be run last. 1. Arrhenius-residual BP model (AR-FHE/residual_bp_model) python prepare_fixed_arrhenius_residual_data.py python train_fixed_arrhenius_residual_bp_final.py --check-data python train_fixed_arrhenius_residual_bp_final.py --profile full --device cpu --cpu-threads 8 2. Direct-stress BP ensemble (BP-E) python train_pure_bp_final.py --check-data python train_pure_bp_final.py --profile full --device cpu --cpu-threads 8 3. Functional GPR/GBR heterogeneous ensemble (AR-FHE) python train_functional_gpr_gbr_ensemble.py --check-data python train_functional_gpr_gbr_ensemble.py --profile full --cpu-threads 8 Before a full run, the end-to-end workflow can be verified with --profile smoke --output-dir outputs_smoke_test. Reproducing the reported predictions without retraining ------------------------------------------------------- The trained checkpoints in final_models/ together with the accompanying manifest files allow the reported predictions to be reproduced by inference alone. The manifests record the selected trials, the structure weights and the per-member ensemble weights. Interruption and resumption --------------------------- Random seeds are fixed in the scripts. Optuna studies, cross-validation results and checkpoints are written incrementally, so an interrupted full run can be resumed by re-issuing the same command. Further detail on the model formulation, validation protocol and output files is given in the instruction files included in each model directory.
Institutions
- Kunming University of Science and TechnologyYunnan, Kunming