The Effectiveness of Machine Learning Models for Newtonian Physics Simulations

Author: Luan A. Tobias
Affiliation: Concordia International School Hanoi / Polygence
Contact: luan.tobias@concordiahanoi.org

🚧 Currently in active development!

🧠 Overview

This repository contains code, data, and results for the research paper:

The Effectiveness of Machine Learning Models for Newtonian Physics Simulations

The goal: Evaluate whether machine learning (ML), specifically neural networks, can simulate Newtonian physics systems (projectile motion and linear drag projectile motion) more efficiently than traditional methods.

📂 Project Structure

• models/: Saved models (kinematics, LDPM)
• notebooks/: Training, analysis, and experiment notebooks (Colab)
• src/: Core simulation and ML model code
• results/: Accuracy curves, benchmarks, and generated figures
• paper/: Drafts and materials for the research paper

⚡ Key Features

• Simulation code for classical projectile motion and LDPM systems in Python/NumPy
• On-the-fly data generation for robust ML training and reduced overfitting
• PyTorch models with various architectures, trained and benchmarked
• Scripts for comparing inference time and accuracy versus analytical solutions
• Metrics implemented: RMSE, MAE, and a custom accuracy formula
• All experiments and findings documented in notebooks and paper drafts

🚀 Quick Start

1. Clone the repository:

   git clone https://github.com/Lu4nAlmeida/cnm-simnet.git
   cd cnm-simnet

2. Install dependencies:

   pip install -r requirements.txt

3. Run simulations or ML training via notebooks in the notebooks/ folder.

📊 Results

See the results/ folder for:

• Accuracy plots
• Inference benchmarks
• Figures for the paper

💡 Research Highlights

• (In Progress)

🤝 Contributing

Interested in helping or using this work? Feel free to reach out or open an issue!

---

“Code is like humor. When you have to explain it, it’s bad.” – Cory House

---

📬 Contact

• Email: luan.tobias@concordiahanoi.org