PolymerMI : Polymer Property Prediction & Optimization (MI/BO/XAI Pipeline)

Purpose

This Pipeline was developed as a lightweight tool that can support laboratory research using only a Linux-based virtual environment. By predicting material properties and performing automated exploration before actual experiments, it helps reduce time and experimental costs. In addition,XAI-based analysis provides interpretability that can guide experiment design and optimization.

SHAP summary plot

Overview This repository provides a lightweight but complete machine learning pipeline for predicting and optimizing polymer properties. It demonstrates core techniques in Materials Informatics (MI), including:

• Synthetic dataset generation
• Random Forest baseline models
• Feature importance analysis (Permutation / Gini)
• Bayesian Optimization (Optuna)
• Gaussian Process Regression (GPR)
• SHAP-based XAI visualization
• Automated workflow (run_all.sh)

Target Property: polymer density (synthetic) Features: mw, hyd, xlink, side, tg_like

Folder Structure

--data_synth.py # synthetic polymer dataset generator

--mi_baseline.py # RF-based MI baseline + feature importance

--bo.optuna.py # Bayesian Optimization (RF)

--bo_gpr.py # Bayesian Optimization (GPR)

--xai_perm_pdp.py # Permutation importance + PDP plots

--xai_shap.py # SHAP summary & dependence plots

--results/ # Figures, CSV logs, best parameters

--run_all.sh # full pipeline runner

Example Outputs

• R² = 0.97 (RF baseline)
• Bayesian Optimization best density = 1.96
• GPR kernel after fitting
• SHAP summary / dependence plots

How to Run bash run_all.sh

Or run individual modules

-python data_synth.py

-python mi_baseline.py

-python bo_optuna.py

-python bo_gpr.py

-python shap_analysis.py

#Environment

-Python 3.10

-Tested in a conda environment (mi310).

Purpose of This Repository This project demonstrates essential workflows in modern Materials Informatics

• nonlinear structure-property modeling

• ML-based polymer property prediction

• ML-assisted polymer design (BO)

• interpretable ML for materials (XAI)

  License MIT License

  #LLZO Ionic Conductivity (Synthetic) — SHAP Analysis#

  

  Target: Ionic conductivity trend sigma(ion) in LLZO (synthetic). Insight: Sintering temperature dominates sigma(ion), followed by dopant fraction and Li excess with clear non-linear effects; dopant type and grain size are secondary.

  

  Dopant fraction was selected for further analysis due to its strong non-linear contribution observed in the global SHAP summary.

  Dependence: The effect of dopant fraction on sigma(ion) is non-linear and modulated by sintering temperature, indicating coupled processing-composition effects.

  How to run

  -python beeswarm.py

  -python shap_LLZO.py

()]