Adsorption Performance Predictor & Adsorbent Selector

This project is a cheminformatics and chemical data science workflow for adsorption-based removal of organic water contaminants (pharmaceuticals, herbicides, fungicides, dyes, etc.).

It combines:

• real experimental adsorption data (~3,700 records)
• molecular representations (RDKit descriptors, fingerprints, ChemBERTa embeddings)
• physics-informed feature engineering
• a hybrid Langmuir + machine learning model
• an interactive Streamlit app for research support

---

🔬 Problem

Given:

• pollutant (structure or descriptors)
• adsorbent properties (biochar composition, surface area, pore structure)
• adsorption conditions (pH, dosage, time, temperature)

Predict:

• adsorption capacity (mg/g)

and support:

• adsorbent design and condition optimization

---

📊 Dataset

Primary dataset: data/raw/ec_biochar_adsorption_raw.csv :contentReference[oaicite:0]{index=0}

• ~3,757 experimental data points
• ~24 variables
• Includes:
  • adsorption conditions
  • biochar properties
  • elemental composition
  • experimental adsorption capacity

⚠️ Real-world characteristics

• heterogeneous sources
• varying experimental conditions
• partial missing data (SMILES coverage ~0.76)

👉 This reflects real industrial chemical data, not curated benchmarks.

---

🧠 Generalizable Molecular Descriptor Framework

A key strength of this project is the use of structure-derived molecular descriptors, including:

• logP (hydrophobicity)
• TPSA (polarity)
• HBD / HBA (hydrogen bonding)
• molecular weight
• ring count

These descriptors are:

✔ computed from structure (SMILES or manual input) ✔ independent of specific datasets ✔ applicable to any organic pollutant

👉 This enables the model and app to:

• evaluate new contaminants not in the dataset
• support exploratory chemical screening
• generalize beyond fixed pollutant lists

---

🧪 Cheminformatics & Data Pipeline

This project implements a modular workflow:

1. Data processing

• cleaning and validation of experimental data
• handling missing and inconsistent values

2. Molecular representation

• RDKit descriptors (interpretable)
• Morgan fingerprints (structural similarity)
• ChemBERTa embeddings (learned representations)

3. Feature engineering

• concentration-to-dosage ratio (C/dosage)
• log transformations
• time-to-dosage interactions

👉 This combines domain knowledge + ML-ready features

---

⚙️ Modeling Approach

Models evaluated:

• Random Forest
• HistGradientBoosting
• SVR (RBF kernel)
• Kernel Ridge

Key finding:

• tree models → step-like, non-physical behavior
• kernel ridge → unstable curvature
• SVR → smooth and physically consistent trends

👉 SVR was selected for hybrid modeling.

---

🔬 Hybrid Langmuir + ML Model

Final model integrates machine learning with adsorption physics:

[ q = \frac{q_{max} K C}{1 + K C} ]

Where:

• ML predicts qmax (capacity limit)
• ML predicts K (adsorption affinity)

Advantages:

• smooth monotonic increase
• realistic saturation behavior
• avoids plateau/decrease artifacts

---

📈 Model Performance

Model	MAE	R²
Hybrid (qmax only)	~10.4	~0.84
Hybrid (qmax + K)	~9.9	~0.88

• strong agreement with experiments
• R² ≈ 0.88
• robust across wide capacity range

---

📊 Key Insights

Feature engineering

• capacity correlates more strongly with C/dosage than concentration alone

Chemical space

• RDKit PCA → physicochemical similarity
• fingerprints → structural similarity
• ChemBERTa → learned relationships

• captures physicochemical properties
• separates molecules by polarity & size

• captures structural similarity
• clusters similar compounds

• learned chemical representation
• captures deeper molecular relationships

Model behavior

• hybrid model preserves physical adsorption trends
• sensitivity analysis shows realistic saturation

• smooth increase
• realistic saturation
• consistent with adsorption physics

---

🧱 Unstructured Data Handling

A prototype parser demonstrates how semi-structured chemical records can be converted into structured features.

👉 This simulates real workflows where data comes from:

• reports
• industrial systems
• regulatory text

---

🚀 Streamlit App: Research Tool

The application is designed as a research support tool, not just a predictor.

Users can:

• input pollutant descriptors (manual or SMILES-based)
• modify adsorbent properties (surface area, composition)
• adjust experimental conditions

Goal:

Explore how changes in biochar properties and conditions affect adsorption capacity

👉 This helps researchers:

• design better adsorbents
• test hypotheses
• guide experiments

---

🧠 Adsorption Recommendation Logic

The app provides chemistry-informed guidance based on:

• hydrophobicity (logP)
• polarity (TPSA)
• aromaticity (π–π interactions)
• surface area and pore structure
• adsorption conditions

👉 Intended as screening support, not a replacement for experiments.

---

🔍 Applicability Domain

Model reliability is constrained by:

• training feature ranges
• chemical space coverage

👉 Input constraints prevent unreliable extrapolation.

---

⚠️ Limitations

• incomplete SMILES coverage (~76%)
• heterogeneous experimental data
• Langmuir assumption may not hold for all systems
• reduced accuracy at extreme values

---

📚 Future Work

• full ChemBERTa integration into model
• pKa / pHpzc incorporation
• multi-component adsorption
• improved SMILES coverage (PubChem mapping)
• integration with chemical databases

---

💼 Relevance for Chemical Data Science

This project demonstrates:

• chemical data pipeline design
• molecular representation (RDKit + embeddings)
• physics-informed feature engineering
• hybrid ML + scientific modeling
• interpretability and domain awareness

---