🧪 Particles Trajectory Prediction

Predicting particle positions from signals detected by Resistive Silicon Detector (RSD) sensors using machine learning.

---

🔍 Project Summary

This project focuses on a regression task: predicting the 2D positions of particles based on signal data collected from RSD sensors. Each event consists of measurements from 18 pads (12 real, 6 noise) and 5 signal features.

---

📦 Dataset

• Total Events: 514,000

• Training Set: 385,500 labeled samples

• Test Set: 128,500 samples

• Features per Event:

  • 5 signal types: pmax, negpmax, tmax, area, rms
  • From 18 pads → 90 features per event

---

⚙️ Preprocessing

• Noise Removal: Excluded 6 noisy pads: 0, 7, 12, 15, 16, 17

• Outlier Filtering:

  • Z-score: Detected abnormal values per event
  • Mahalanobis Distance: Accounted for feature correlations

• Feature Selection:

  • Used Random Forest to assess feature importance
  • Kept: pmax, negpmax, area

---

🤖 Models

Random Forest Regressor ✅

• Best performance
• Handles noise and high-dimensional data well

k-Nearest Neighbors

• Local-based predictions
• Slower and less robust to noise

---

🔧 Hyperparameter Tuning

Used grid search to optimize:

Random Forest

• n_estimators: 300
• max_depth: 40
• max_features: sqrt
• min_samples_leaf: 2
• min_samples_split: 2
• criterion: squared_error

kNN

• n_neighbors: 10
• weights: uniform
• algorithm: auto

---

📊 Results

Model	R² Score	MAE	MSE	Public Score
Random Forest	0.9991	2.67	12.16	4.750
kNN	0.9985	2.81	20.12	5.683

Public Score: Avg. Euclidean distance between predicted and true positions

---

📈 Key Takeaways

• Successfully filtered sensor noise and irrelevant features
• Random Forest yielded highly accurate particle position predictions
• Achieved R² score of 0.9991 on validation
• Scalable approach with potential for deep learning extensions

---

🧠 Future Work

• Experiment with neural networks for generalization
• Refine hyperparameter tuning with advanced search techniques
• Add k-fold cross-validation for robust evaluation
• Explore additional denoising and signal enhancement methods

---

👥 Authors

• Tommaso Mazzarini – Politecnico di Torino – tommaso.mazzarini@studenti.polito.it
• Leonardo Merelli – Politecnico di Torino – leonardo.merelli@studenti.polito.it