kkt-hardnet

kkt-hardnet is the publishable Python package for KKT-HardNet.

Install

Editable install from this repository:

pip install -e kkthn

Editable install with CUDA 12 support:

pip install -e "kkthn[cuda12]"

Or use pip after publishing:

pip install kkt-hardnet

pip install "kkt-hardnet[cuda12]"

Import

from kkthn import KKTHardNet

Core Methods

• model() for supervised surrogate learning from parameters.csv and variables.csv
• optimize() for unsupervised optimization from parameters.csv
• estimate() for inverse parameter estimation from parameters.csv and variables.csv
• load(metadata_path) to reload a trained run
• predict(x) to infer variables for new parameter values

Packaging

The PyPI project name is kkt-hardnet, while the import path remains kkthn. The default install is CPU-oriented; CUDA support is selected explicitly with the cuda12 extra instead of automatic device detection.

Modeling Workflow

Use KKTHardNet to define symbolic constrained problems with:

• named parameters, decision variables, and optional inverse parameters
• equality and inequality constraints written as Python expressions
• optional objectives for optimization tasks
• hard KKT projection during training and inference
• saved run artifacts including weights, predictions, history, summary, and metadata

Available methods:

• dataset(parameters=..., variables=...) to attach training data
• model() for supervised surrogate learning
• optimize() for unsupervised optimization
• estimate() for inverse parameter estimation
• load(metadata_path) to reload a trained run
• predict(x) to evaluate a trained or loaded model on new inputs

The example below shows one complete workflow. The CSV column names must match the declared parameter and variable names. Each run writes <model_name>_<timestamp>/ with parameters.csv, optional variables.csv, history.csv, predictions.csv, model_weights.npz, summary.json, and metadata.json.

from kkthn import KKTHardNet

TRAIN = {
    "epochs": 1200,
    "batch_size": 32,
    "learning_rate": 1e-3,
    "train_frac": 0.8,
    "hidden_size": 64,
    "hidden_layers": 2,
    "seed": 42,
    "dtype": "float64",
    "print_every": 1,
    "newton_step_length": 0.5,
    "newton_tol": 1e-6,
    "newton_reg_factor": 1e-3,
    "max_newton_iter": 30,
    "max_backtrack_iter": 10,
}

# Build a symbolic problem.
model = KKTHardNet(name="demo_model", train=TRAIN)
x = model.add_parameter(["x1", "x2"])
theta = model.add_inverse_parameter(["a0", "a1"], init_value=[10.0, -10.0])
y = model.add_variable(["y1", "y2", "y3"])

# Objectives are optional for surrogate modeling and inverse estimation,
# but required for optimize().
model.objective = 0.5 * (y.y1**2 + y.y2**2 + y.y3**2)
model.constraints.add(
    theta.a0 * y.y1 + y.y2 - x.x1 == 0,
    y.y2 - theta.a1 * y.y3 - x.x2 == 0,
    y.y1**2 + y.y3**2 <= 2.0,
    y.y1 >= 0,
)

# Attach data.
# For model() or estimate(): provide both parameters and variables.
# For optimize(): provide only parameters.
model.dataset(parameters="parameters.csv", variables="variables.csv")

# Choose one training mode.
surrogate_result = model.model()
# optimize_result = model.optimize()
# estimate_result = model.estimate()

# Reload a saved run later and predict on a new parameter vector.
reloaded = KKTHardNet()
reloaded.load("demo_model_20260414_120000/metadata.json")
prediction = reloaded.predict([1.0, 2.0])

⚠️ Please cite our work if you use this code in your research. Citation formats are provided below.

arXiv Preprint: https://arxiv.org/pdf/2507.08124 Journal: https://doi.org/10.1016/j.compchemeng.2025.109418

@article{iftakher2025physics,
  title={Physics-informed neural networks with hard nonlinear equality and inequality constraints},
  author={Iftakher, Ashfaq and Golder, Rahul and Nath Roy, Bimol and Hasan, MM Faruque},
  journal={Computers \& Chemical Engineering},
  pages={109418},
  year={2025},
  publisher={Elsevier}
}