cavity-encoding-ml

Cavity encoding and leakage-safe validation for machine learning on multi-cavity injection molding data.

A multi-cavity mold forms several parts in every machine cycle (a shot). When you predict part quality from process data, three questions come up.

1. Do the cavities differ enough that the cavity label should enter the model?
2. How should the label be encoded?
3. How do you validate the model when all parts of a shot share the same machine settings?

cavity-encoding-ml answers them in four steps. It implements the method of Boakye, Hafizh, Lohar and Abdul Majeed (accepted, Journal of Intelligent Manufacturing, 2026) and works with any machine learning library that offers fit and predict.

Install

pip install cavity-encoding-ml

Before the release is on PyPI you can install from GitHub:

pip install "git+https://github.com/nanaboak/cavity-encoding-ml.git"

Python 3.9 or newer. The core needs only NumPy, pandas, SciPy and scikit-learn. Figures need pip install matplotlib.

Try it in one line

cavity-encoding-ml example

This runs the four steps on synthetic tip diameter data that is shipped with the package.

Quick start

from sklearn.ensemble import RandomForestRegressor
from cavity_encoding_ml import CavityConfig, CavityWorkflow
import pandas as pd

df = pd.read_csv("my_parts.csv")          # one row per part

config = CavityConfig(
    shot="ShotID",                         # column that identifies the machine cycle
    cavity="CavityNumber",                 # column with the cavity number of the part
    response="TipDiameter",                # column with the measured characteristic
    shared=["InjPressure", "CoolFlow"],    # machine features, equal for all parts of a shot
    sensors={                              # one cavity sensor type, one column per cavity
        "Max": ["Cav1Max", "Cav2Max", "Cav3Max", "Cav4Max"],
    },
    cavities=[1, 2, 3, 4],
)

result = CavityWorkflow(config, RandomForestRegressor(random_state=0)).run(df)
print(result.report())
print(result.verdict)                      # "use_cavity", "adds_noise" or "no_cavity_effect"

The workflow trains the model twice on identical shot-grouped folds, without and with the cavity label, and then decides. Add arms=("blind", "readings", "encoded") to also see the cavity readings without the label (R1), and audit_kfold=True to also measure how much ordinary K-fold would flatter the result. Both are optional and cost extra training time.

If your data have no cavity sensors, leave sensors empty. The package then compares the model with and without the cavity label alone.

Only train a model with the cavity label

If you only want the library to build the features and cross-validate one model, without any comparison:

from cavity_encoding_ml import validate_cavity_encoding

v = validate_cavity_encoding(df, config, RandomForestRegressor(random_state=0),
                             encoding="onehot", arms=("encoded",))
print(v.encoded.rmse, v.encoded.r2)        # one model, shot-grouped folds, leakage check included

Any machine learning library

from cavity_encoding_ml import validate_cavity_encoding, FunctionModel

validate_cavity_encoding(df, config, XGBRegressor())          # any object with fit and predict
validate_cavity_encoding(df, config, lambda: LGBMRegressor()) # or a function that returns one
validate_cavity_encoding(df, config, FunctionModel(my_fit, my_predict))  # PyTorch, TensorFlow, JAX

What is in this repository

Verification

python scripts/verify_paper3_pipeline.py compares the package with a literal copy of the fold and representation code of the original study. For the random forest the pooled RMSE is exactly equal. For gradient boosting, support vector regression and the multilayer perceptron it agrees to machine precision. On the confidential tip diameter and tensile strength data, the F test reproduces F(3, 285) = 326.4 and F(3, 1513) = 11.4, and the decision matches the original study. The production data are not part of this repository. scripts/reproduce_paper3.py repeats the analysis if you have the files.

Cite

Use the button Cite this repository on GitHub, which reads CITATION.cff, and cite the original study.

Contribute and license

See CONTRIBUTING.md and CODE_OF_CONDUCT.md. The package is released under the MIT license.