PGML Copper Heap Leaching

Physics-Guided Machine Learning for Sustainable Heap Leaching Control
Department of Chemical Engineering, University College London
Authors: Sultan Alhamdan, Dr. Paulina Quintanilla (p.quintanilla@ucl.ac.uk)

Source code for a physics-guided machine learning model that predicts and regulates bed saturation in copper heap leaching. It replaces the per-step hydraulic re-fitting a Richards-based nonlinear model predictive controller needs, moving identification into training so one controller can run at heap scale. Baseline: Olivares et al., 2025.

The hydraulic transport layer uses two constitutive relations:

• van Genuchten Modified (VGM) drives data generation
• Gardner drives the controller and the physics penalty during training

The network is a 4-layer MLP of width 512, trained on a convex combination of a data term and a Gardner physics residual at λ = 0.4. Input is 226 values: the 221-node saturation profile, the irrigation rate over the step, and four soil properties. Output is the 221-node profile one hour ahead.

1. Download dataset.h5 and put it in src/data/. See src/data/README.md.

2. Install dependencies. torch is a nightly build and is not on PyPI, so it installs separately:

   pip install torch==2.12.0.dev20260408+cu128 --index-url https://download.pytorch.org/whl/nightly/cu128
   pip install -r requirements.txt
   

cupy is optional. Without it the solver runs on the CPU, correctly but slowly.

No configuration. Every script finds data/ and reference/ through paths.py when it starts. Keeping the dataset somewhere else is covered in src/data/README.md.

cd src/06_Control
python closed_loop.py                    # 614 validation soils
python closed_loop.py --split test       # 615 test soils

Uses the trained weights in 04_Training/models/ and the inputs in reference/. Writes to results/closed_loop/<split>/.

cd src/03_Preprocessing && python preprocess.py
cd ../04_Training       && python train.py --lam 0.4 --seed 42 --out runs/s42_lam0p4
cd ../04_Training       && python summarise_training.py --root runs
cd ../06_Control        && python closed_loop.py --sweep_dir ../04_Training/runs

preprocess.py overwrites reference/scaler.json and reference/split.json with identical values, since the split seed is fixed.

src/
    data/                 download dataset.h5 and save it here
    reference/            inputs the code reads: scalers, splits, soil lists, ceilings
    results/              records from the released runs
    paths.py              resolves every path

    01_Simulator/
        column_solver.py           Richards solver, batched over columns
        tridiagonal_solve.py       Thomas algorithm as a CUDA kernel
    02_Data_Generator/
        generate_dataset.py        4096 Sobol to create dataset.h5
    03_Preprocessing/
        preprocess.py              split, scaling, per-timestep samples
    04_Training/
        train.py                   training loop
        physics_loss.py            Gardner residual
        fit_alpha_g.py             derives the Gardner alpha_g rule
        summarise_training.py      reduces run logs to one summary
        models/                    three checkpoints, λ 0.4, seeds 7/42/123
    05_Evaluation/
        openloop_drift.py          drift when the network runs unanchored
    06_Control/
        closed_loop.py             the controller
        rmax_ceiling.py            per-soil irrigation ceiling
        speed_benchmark.py         times the solver on the CPU
        extrapolation/
            build_extrapolation_soils.py    soils outside the training domain

Run any script with --help to see its options.

Three trained checkpoints are included here, at λ = 0.4 and seeds 7, 42 and 123.

Dataset is hosted on OneDrive:

• dataset.h5, 4096 simulated leach cycles

Download: OneDrive