EMI Surrogate Modeling with SiC MOSFETs

This repository contains the complete pipeline for building surrogate machine learning models to predict Electromagnetic Interference (EMI) metrics from SiC MOSFET datasheet parameters and LTSpice simulation data.

The project forms the basis of a dissertation that explores generalization across multiple MOSFET devices, using both classical machine learning and deep learning models.

Motivation

Wide-bandgap (WBG) devices such as Silicon Carbide (SiC) and Gallium Nitride (GaN) MOSFETs enable compact, efficient, and high-performance power converters for electric vehicles. However, their rapid switching transitions introduce steep dv/dt and di/dt, which interact with parasitic inductances and capacitances to cause overshoot, undershoot, ringing, and electromagnetic interference (EMI).

Traditional EMI evaluation relies on double-pulse test (DPT) simulations in LTSpice with MATLAB post-processing. While accurate, this approach is computationally heavy and slow, especially when scaling to multiple devices, parameter sweeps, and operating conditions. This creates a bottleneck for fast design iteration and delays EMI-aware optimisation.

Surrogate machine learning models offer a scalable alternative by directly learning the nonlinear mapping between device parameters and EMI targets from simulation data. Once trained, these models provide near real-time EMI predictions, generalise to unseen devices, and support systematic exploration of best- and worst-case switching scenarios. Artificial Neural Networks (ANNs) in particular are well-suited for capturing the high-dimensional nonlinear behaviours of switching transitions.

---

Research Aims, Novelty, and Contributions

• Artificial Neural Networks (ANNs) for Surrogate EMI Prediction

  • Developed a multi-output ANN capable of predicting 13 switching waveform characteristics (rise/fall times, overshoot, undershoot, ringing).
  • Achieved strong generalisation performance on unseen MOSFET devices.

• Benchmarking with Classical Baselines

  • Implemented Ridge, SVR, Random Forest, LightGBM, and XGBoost for multi-output regression, highlighting limitations in device generalisation.

• Iterative ANN Architectural Development

  • Progressed from Baseline MLPs → Deep MLPs → Multi-Head → Masked → Attention-based ANNs.
  • Incorporated dropout, L2 regularisation, batch normalisation, early stopping, learning-rate scheduling, physics-informed features, and per-target scaling.

• Classification of Switching Behaviour

  • Extended surrogate modelling to identify best/neutral/worst-case operating conditions, supporting EMI-aware design.

• Harmonic Spectrum Analysis and THD

  • Computed harmonic spectra and Total Harmonic Distortion (THD) as indicators of EMI severity, linking waveform modelling to spectral EMI metrics.

---

Objectives

1. Extract and preprocess simulation and datasheet parameters for six SiC MOSFETs.
2. Perform exploratory data analysis (EDA) to understand input features and EMI target characteristics.
3. Build a clean, balanced dataset with consistent simulation setups across devices.
4. Implement feature engineering with physics-informed derived features.
5. Train and evaluate classical ML models (Ridge, SVR, Random Forest, LightGBM, XGBoost).
6. Train and optimize deep learning models (ANN, attention-based, multi-headed architectures).
7. Evaluate generalization performance on unseen MOSFET devices.
8. Deploy the best surrogate model for EMI prediction and scenario analysis.

---

Repository Structure

Please check the project_structure.txt tree /f > project_structure.txt

PHASE 1: SWITCHING WAVEFORM WAVEFORM SURROGRATE MODELLING

Workflow Overview

1. EDA

   • EDA_Features_Understanding.ipynb
   • input_summary_per_mosfet.ipynb

2. Data Preprocessing

   • Extraction:
     • extract_sic_mosfet_parameters.ipynb
     • extract_tables_from_datasheets_pdf.ipynb
   • Merging & Cleaning:
     • Merging_Simulation_and_MOSFET_data.ipynb
     • Outlier_Negatives_and_Null_values.ipynb
     • Outlier_Removal_Each_MOSFET/
     • Merged_Final_Complete.ipynb
   • Feature Engineering:
     • Feature_engineering.ipynb
     • feature_engineering.py

3. Classical Models

   • Ridge Regression → Ridge_Regression.ipynb
   • Support Vector Regression (SVR) → SVR.ipynb
   • Random Forest → RandomForest.ipynb
   • LightGBM → LightGBM.ipynb
   • XGBoost → XGBoost.ipynb

   Balanced FAST sampling for quick iteration

   • 70/15/15 split (via 70/30 then halved)
   • Scale outputs ONLY (inputs left in physical units)
   • Metrics (R2, RMSE, MAE) per target + overall
   • Plots: Val vs Test R² bars, Test scatter, Residual histograms

4. Neural Network Models - ANN

• Multiple ANN architectures developed for EMI surrogate modeling:

  • Baseline ANN (BaselineANN.ipynb) – stepwise improvements (regularization, batch norm, dropout, schedulers).
  • Attention ANN (Attention_ANN.ipynb) – integrates attention layers to focus on key input features.
  • Deep MLP (Deep_MLP.ipynb) – deeper multilayer perceptron to capture complex nonlinearities.
  • Masked ANN (Masked_ANN.ipynb) – applies feature masking per output target.
  • Multi-Head ANN (Multi_head_ANN.ipynb) – independent branches per EMI output for multi-task learning.

• Common outputs across models:

  • Trained .h5 models
  • r2_rmse_tables/ (train, val, test, unseen metrics)
  • predicted_vs_actual/ scatter plots
  • Training/validation loss curves

These ANN models extend the classical baselines, focusing on non-linear feature learning, generalization to unseen MOSFETs, and architectural experimentation to maximize EMI prediction accuracy.

5. Final Model: Multi-Head ANN + Light Attention

Folder: FINAL_MODEL/final_multihead_attention/
Notebook: FINAL_MODEL/FINAL_MODEL.ipynb

• Description: Final surrogate model predicting 13 EMI waveform parameters.
• Key Features: Multi-head ANN, light attention, physics-informed inputs, generalisation to unseen MOSFETs.
• Outputs:
  • models/final_multihead_attention.h5 – trained Keras model
  • train_val_loss_curves/loss.png – training vs validation loss
  • r2_rmse_tables/ – R² & RMSE tables (train.csv, val.csv, test_seen.csv, unseen.csv)
  • predicted_vs_actual/test_seen_scatter.png – scatter plots
  • predicted_vs_actual/test_seen_inputs_true_pred.csv – inputs + true vs predicted
  • artifacts/ – scalers & column configs for inference

6.Switching Scenarios: Physics-Informed Severity Classification

Folder: Switching_Scenarios_Physics/
Notebook: Switching_Scenarios_Physics/Classifier.ipynb
Docs: Switching_Scenarios_Physics/README_Switching_Severity_Classification.md

• Description: Classifies switching scenarios as Best, Neutral, or Worst using physics-informed scoring (rise/fall times, overshoot/undershoot, ringing frequency).
• Per MOSFET Outputs:
  • {device}_labeled.csv – original data + z-scores + severity score + scenario label
  • {device}_top5_best.csv / {device}_top5_worst.csv – top-5 switching cases
  • {device}_score_hist.png – score distribution with thresholds
  • {device}_PCA_best_worst.png – PCA scatter of switching behaviour
  • {device}_contribution_bar.png – parameter contributions (Best vs Worst)
  • {device}_contrib_rows.csv – per-row contribution breakdown

7. Full Data Run: Multi-Head ANN

Folder: Full_Data_Run_MultiHead_ANN/
Notebook: Full_Data_Run_MultiHead_ANN/COMPLETE_RUN.ipynb

• Description: Final complete run of the multi-head ANN on the entire dataset, saving full evaluation tables, loss curves, predicted vs actual plots, and trained models.

---

Outputs

• Clean balanced datasets

  • All_6_MOSFETs.csv → Complete merged dataset
  • Train_5_MOSFETs.csv → Training set (5 devices)
  • Test_1_MOSFET.csv → Held-out unseen MOSFET

• EDA reports

  • Feature distributions
  • Simulation input summaries per device
  • Feature–target correlations

---

PHASE 2: HARMONIC SPECTRUM ANALYSIS AND TOTAL HARMONIC DISTORTION (THD)

• Module: harmonics_pipeline.py (see README_Harmonics.md)
• Input: Harmonics_Data/harmonics_*.csv with Frequency_k_Hz, Mag_k, Phase_k per row
• Steps: IQR outlier removal → THD% → dB spectrum → noise floor → decay slope → band powers (100 kHz–3 MHz, 3–10 MHz, 10–30 MHz) → peak count/max above floor → risk flags
• Outputs:
  • Per-file features: Harmonics_Data/processed/<file>_features.csv
  • Master table: Harmonics_Data/processed/emi_master_features.csv
  • IQR summary: Harmonics_Data/processed/iqr_summary.csv
• Plots (notebooks): harmonic components, composite waveform, spectrum bars

Notes

• All preprocessing ensures consistent simulation setups across six MOSFET devices.
• Final evaluation prioritizes generalization to unseen devices, which is critical for practical EMI modeling.
• The repository is modular: EDA, preprocessing, feature engineering, classical ML, and ANN pipelines are separated for clarity.

---

References

• SiC MOSFET datasheets (Wolfspeed)
• LTSpice simulation datasets
• EMI and power electronics literature