🧪 Experimental and Machine Learning-Based Prediction of Properties in Bi Off-Stoichiometric NBT–7BT Ceramics

This repository contains the complete experimental + machine learning (ML) + deep learning (DL) workflow for studying and predicting the structural, dielectric, ferroelectric, and piezoelectric properties of Bi off-stoichiometric NBT–7BT ceramics.

Author: Srishty Mangutte
Department: Ceramic Engineering, NIT Rourkela
Guide: Prof. Anupam Mishra

---

📌 Project Overview

Lead-free piezoelectric ceramics such as NBT–BT (Na₀.₅Bi₀.₅TiO₃–BaTiO₃) are promising alternatives to lead-based materials. However, Bi volatility during high-temperature sintering leads to A-site off-stoichiometry, which strongly affects:

• Crystal structure
• Microstructure
• Dielectric behavior
• Ferroelectric response
• Piezoelectric performance

To reduce experimental trial-and-error, this project integrates:

• Experimental electroceramics characterization
• Machine Learning for composition → property prediction
• Deep Learning for SEM microstructure → grain size prediction

---

🏗️ System Architecture

Pipeline:

1. Experimental synthesis of Bi-deficient & stoichiometric NBT–7BT
2. Structural & electrical characterization
3. Data extraction from plots and SEM images
4. ML model training for property prediction
5. DL model training for microstructure-based grain size prediction

---

🧫 Experimental Work

The following experimental techniques are used:

• XRD → Phase identification and structural analysis
• SEM → Microstructure & grain size analysis
• Archimedes Method → Density & porosity
• LCR Meter → Dielectric measurements
• Ferroelectric Loop Tracer → P–E hysteresis loops
• Strain Measurement System → S–E butterfly loops
• Piezometer → d₃₃ measurement

Images here:

---

🤖 Machine Learning Model – Property Prediction

Objective

Predict the following properties:

• d₃₃
• Dielectric constant (εᵣ)
• Coercive field (E𝒸)
• Grain size

Input Features

• SEM images
• Bi off-stoichiometry (Biₓ)
• Ba content
• Sintering temperature
• Bulk density

Model Details

• Algorithm: Random Forest Regressor
• Framework: Scikit-learn
• Data Split: 80% Training / 20% Testing
• Evaluation Metrics: R², MAE, RMSE

ML output plots:

---

🧠 Deep Learning Model – SEM Image → Grain Size

Objective

Predict average grain size directly from SEM microstructure images.

Dataset

• 2–3 SEM images per experimental sample
• Grain size labels extracted from Excel data (mean of Length column)

Model Details

• Framework: TensorFlow (Keras API)
• Architecture: ResNet-50
• Training: Transfer learning with fine-tuning
• Optimizer: Adam
• Loss Function: Mean Squared Error (MSE)

DL images:

---

⚙️ Setup Instructions

git clone https://github.com/Srish-ty/resnet-SEM-microstructure-predictor
cd resnet-SEM-microstructure-predictor
pip install -r requirements.txt

Run DL notebook:

jupyter notebook notebooks/dl_sem_grain_size_prediction.ipynb

📊 Key Outcomes

• ML successfully predicts electromechanical properties from composition.

• DL model estimates grain size directly from SEM es.

• Integration of ML + DL reduces experimental trial cycles.

• Demonstrates data-driven electroceramics design.

📚 Data Source & Credibility

• Experimental data obtained from NBT–7BT Bi off-stoichiometric ceramic synthesis and characterization under Prof. Anupam Mishra.

• Grain size labels extracted from experimental SEM e analysis.

• ML datasets constructed using:

• Experimental trends

• Peer-reviewed electroceramics literature ranges

• ML and DL models are used as proof-of-concept predictive tools.