Overview of Classification+Regression Approach to predict Best Solver and Best Solver Time

Pipeline Overview

1. Step 1 — Run inside the Data_to_D_with_fractures.ipynb in PorePy environment
   Go to: https://github.com/keileg/solver_configuration_flow_setups/blob/main/Data_to_D_with_fractures.ipynb
   Input: Model setup and fracture configuration.
   Output:

   • Mesh files: mesh2d_Xfr_id.msh
   • Fracture metadata: mesh_meta_id.json
   • Solver log: solver_times.csv (convergence, solve time, etc.)

   Example:
   .../Data_Large4 contains 4000 meshes, matching fracture files with IDs in 0.4d format, and one solver_times.csv.

---

2. Step 2 — Create the feature dataset Input:

   • Mesh files *.msh
   • Fracture metadata *.json
   • Solver log solver_times.csv

   Output:

   • Feature table mesh_features.csv combining mesh geometry, fracture stats, and solver outcomes.

   Usage example:

   from dataset_builder import MeshFeatureDatasetBuilder
   
   builder = MeshFeatureDatasetBuilder(
       mesh_dir="D:/ML4pmg/Data_Large4/",
       solver_csv="D:/ML4pmg/Data_Large4/solver_times.csv",
       output_csv="D:/ML4pmg/Data_Large4/mesh_features.csv",
       log_transform=True
   )

---

3. Step 3 – Run Clf_and_Reg_Combined.ipynb
   • Train models and generate predictions.
   • Compare results against baseline and clairvoyant approaches.