top of page
Search

From First-Year Civil to SciML Pioneer: How I Neural ODEs Epidemic Modeling with AI & Julia! 🚀

  • Writer: Vveer Motani
    Vveer Motani
  • May 29, 2025
  • 2 min read



I'm diving headfirst into Scientific Machine Learning (SciML)! It's incredible how SciML tackles complex, real-world problems traditional methods can't. My latest project, "Building Neural ODEs for the SIR Model," was a game-changer. 📈


📌 The Hard Part: Modeling the Unknown How do you predict disease spread or other dynamic systems when the underlying equations are unclear or data is incomplete? That's the real challenge. My work in Neural ODEs Epidemic Modeling directly addresses this. 🤔


💡 My Secret Weapon: Neural ODEs + Julia! I used Neural Ordinary Differential Equations (Neural ODEs) to teach an AI to learn epidemic dynamics directly from data! Instead of defining complex equations, my neural network learned the time-evolution itself. ✨

Julia made this possible! Its speed, expressive syntax, and the SciML ecosystem (DifferentialEquations.jl, DiffEqFlux.jl) were a lifesaver. Julia truly came to the rescue! 🧠


🛠️ How I Built It:

  • Generated synthetic SIR data.

  • Implemented a Neural ODE to learn system dynamics.

  • Trained with ADAM optimizer (0.05, 0.01 learning rates) over 100-1000 iterations.


📉 Key Learnings:

  • Stability vs. Speed: Lower learning rates (0.05) gave smoother, more accurate training. Higher rates (0.1) were faster but volatile.

  • Deeper Nets: More complex models learned faster in fewer iterations, despite slower training.


⚙️ My Roadblocks:

  • High initial loss, but consistent decrease.

  • Non-monotonic loss (spikes then settled).

  • Couldn't get GPU support working. 😫

  • Minor post-training script error (didn't affect results).

🌍 Why This Matters: Neural ODEs Epidemic Modeling is crucial for inferring dynamics from partial observations, vital for real-world scenarios.


🧭 Next Steps: Real-world epidemic data, uncertainty modeling, GPU acceleration!

This is just the start of my journey bridging AI and classical science. Huge thanks to Raj Abhijit Dandekar for his incredible mentorship in SciML with Julia! 🙏




Comments

bottom of page