ML workflows for screening degradation-relevant properties of forever chemicals (PFAS)

Physics-informed machine learning workflows for rapid, high-throughput prediction of the molecular properties that govern PFAS degradation, so that candidate molecules and bonding motifs can be prioritized before expensive DFT or experimental remediation studies.

Reference implementation for:

P. Ray, A. R. Castillo, M. Kolel-Veetil, S. R. Kalidindi, "ML Workflows for Screening Degradation-Relevant Properties of Forever Chemicals," Advanced Science (2026): e23817. https://doi.org/10.1002/advs.202523817

Open-access full text: Wiley · PMC · PubMed

---

Motivation

The environmental persistence of per- and polyfluoroalkyl substances (PFAS) demands new remediation technologies, but the PFAS chemical space is far too large to explore with DFT or experiment alone. Rational design of degradation strategies hinges on three molecular properties:

• Bond dissociation energies (BDEs) — govern degradation kinetics
• Polarizability — controls catalytic interactions
• Thermodynamic stability (enthalpy / free energy) — governs reaction feasibility

The key modeling insight is that these properties are not all governed by the same information: global properties (polarizability, stability) depend on spatially-informed 3D electron-density patterns, while bond-specific properties (BDE) are governed by topological atomic connectivity. Each property class therefore gets a featurization matched to its underlying physics.

Two workflows

1. Global properties — voxelized spatial statistics + GPR Atomic attributes (e.g., first ionization energy, electronegativity, and their products) are mapped onto a 3D voxelized grid. Two-point spatial correlations produce a high-dimensional structural fingerprint encoding 3D conformation and chain geometry, which is compressed via PCA (retaining >95% of structural variance) into a compact feature vector. This feeds a Gaussian Process Regression (GPR) surrogate that also returns calibrated uncertainty.

2. Bond-specific properties — graph featurization + Random Forest Bond-level BDEs are predicted from topological / graph-based features (atomic connectivity) using a Random Forest (RF) model.

Key results

Property	Feature engineering	Model	Performance
Polarizability	Voxelized 2-pt correlations → PCA	GPR	R² ≈ 0.92
Enthalpy (ΔH) / stability	Voxelized 2-pt correlations → PCA	GPR	R² ≈ 0.97
Bond dissociation energy	Graph / topological	RF	R² ≈ 0.87

Repository structure

screenPFAS/
├── mol_analysis/            # Core package: cheminformatics + ML for PFAS
├── nbs/wong_analysis/       # Supporting analysis notebooks
├── paper_data/              # Data/artifacts backing the paper figures
├── docs/                    # Documentation (nbdev)
├── 00_core.ipynb            # Source notebook (nbdev-generated module)
├── index.ipynb              # Package index / landing notebook
├── screenPFASworkflow.ipynb # Tutorial: end-to-end workflow
├── pfas_environment.yml     # Conda environment
├── docker-compose.yml       # Containerized setup
├── settings.ini, setup.py, Makefile, MANIFEST.in  # nbdev / packaging
├── CONTRIBUTING.md
└── LICENSE                  # MIT

The package is built with nbdev: notebooks are the source of truth and the mol_analysis module is generated from them.

mol_analysis modules

• descriptors — transform molecular data into ML-ready representations (voxelization, spatial correlations, graph features)
• visualizations — visualize molecules and molecular descriptors
• transforms — transform molecules and molecular descriptors
• testdata — bundled test data (DFT)
• models — testbed ML models (GPR, Random Forest)

Installation

Via pip:

pip install mol_analysis

Or from source (recommended for reproducing the paper):

git clone https://github.com/pranoy-ray/screenPFAS.git
cd screenPFAS
conda env create -f pfas_environment.yml
conda activate pfas        # env name per pfas_environment.yml
pip install -e .

A containerized setup is also provided:

docker compose up

Quickstart

Open the tutorial notebook for the full pipeline (featurization → training → uncertainty → SHAP → figures):

jupyter lab screenPFASworkflow.ipynb

Reproducibility

Scripts for feature engineering (voxelization and graph-based featurization), model training (GPR and Random Forest), and post-processing (uncertainty analysis, SHAP calculations, and figure generation) are contained in this repository.

The curated PFAS property datasets are derived from ERDC and EPA sources and are available from the authors on reasonable request, subject to the data owners' institutional data-release policies.

Citation

@article{ray2026pfasworkflows,
  title   = {ML Workflows for Screening Degradation-Relevant Properties of Forever Chemicals},
  author  = {Ray, Pranoy and Castillo, Andrew R. and Kolel-Veetil, Manoj and Kalidindi, Surya R.},
  journal = {Advanced Science},
  year    = {2026},
  pages   = {e23817},
  doi     = {10.1002/advs.202523817}
}

License

Released under the MIT License.

Contact

Corresponding author: Surya R. Kalidindi, Georgia Institute of Technology. Code author: Pranoy Ray — pranoy-ray.github.io · GitHub @pranoy-ray