Why SciML is the best way to transition to ML from any traditional domain?

Learn about our live upcoming course.

Imagine you are a mechanical engineer or economist or astrophysicist. You would like to capitalize on AI and incorporate it into your work.

Firstly, why is this helpful? Because there are domain experts and there are ML expert. But how often do you find a person who is very good at say structural engineering + machine learning? This is where Scientific Machine Learning enters the picture.

The problem with traditional way of learning ML

This is what most people do.

• Step 1: Learn ML Theory

• Step 2: Do toy Kaggle projects

• Step 3: Update your resume claiming you are good at ML

I don’t have any problem with Step 1. It is a sensible thing to do. But Step 2 and Step 3 are the problems.

Projects done in traditional ML courses

Tell me how are these projects relevant to say a mechanical/electrical/civil engineer?

1. Automatic movie analyzer

2. MNIST digit recognizer

3. Movie recommendation systems

4. Titanic surviver prediction

These can be called as “Toy Kaggle Projects”.

You can play with toy Kaggle projects. But if you want to compete, you have to do serious ML projects. It was good to work on these projects to show expertise in 2012-2014. But not anymore!

We need the following kind of projects

Have a look at the below projects. They are serious ML projects that combine domain specific knowledge with ML. These are research papers published and accepted at top journals/conferences.

If you learn to do these projects on your own…

Is ML only meant for CS students? Absolutely not.

• I had no background in formal CS

• I did not take Data Structures, Algorithms courses

• I did not take any CS course in undergraduate

• I was a core mechanical engineer

ML is becoming more and more democratized. You don't need a formal CS degree to have a career in ML.

Scientific Machine Learning: Domains who can transition

Here are some domains from where you will find SciML engineers and scientists.

• Mechanical

• Electrical

• Chemical

• Biological

• Physics

• Economics

• Data scientists

• Geologists

• Nuclear engineers

• CFD/Fluid mechanics

• Computer Science

• Civil

• Finance

• Urban planning

• Battery engineer

• Bioengineer

• Automation

• Acturial Science

• Robotics

Introduction to Scientific ML

Scientific Machine Learning (SciML) is the branch of ML which combines ML techniques with domain knowledge. In SciML, we can combine the power of neural networks with the interpretability of scientific structures like differential equations.

Any system that can be represented using an Ordinary or Partial Differential Equation is a potential candidate to setup Scientific ML problem.

Simple neural networks have the potential to learn wide variety of functions. Refer to Universal Approximation Theorem.

We combine the power of mechanistic models to be interpretable and the power of neural networks to be universal function approximators in SciML.

A bit about me

Starting from this January 23rd, 2025, I will be teaching Scientific ML live through our online course. If you are interested to join you can register here. We will close the registrations in a few days: Registration link.

Here is what we will be covering in the course

1. Course introduction

   • Transitioning to ML

   • Introduction to SciML

2. What is Julia?

   • Installing and understanding Julia programming language

   • Basic Julia programming

3. Differential equations

   • Introduction to differential equations

   • ODEs

   • PDEs

4. Differential equations in Julia

   • Running ODEs in Julia

   • Running PDEs in Julia

   • Solving ODEs in Julia

   • Solving PDEs in Julia

5. Introduction to Neural Networks

   • Feedforward neural networks

   • Gradient descent introduction

   • Backpropagation

6. The 3 pillars of Scientific ML

   • PINNs

     • Physics Informed Neural Networks (PINNs) - theory

     • Physics Informed Neural Networks (PINNs) - practical

   • Neural ODEs

     • Neural ODE - theory

     • Neural ODE - practical

   • Universal Differential Equation

     • UDE - theory

     • UDE - practical

   • Specific examples from each categories

     • 1D wave equation simulation using PINN

     • Building Neural ODE in Julia

     • Modeling UDE in Julia

7. Hands-on research projects

   • Covid-19 infection prediction using SciML

   • Black hole dynamics using SciML

   • Modeling EV battery degradation using SciML

   • Chandrasekhar white dwarf equation modeling using SciML

8. How to write a research paper?

   • Idea to results

   • Results to story

   • Story to paper

9. Next step: SciML research

   • Potential topics

   • Top journals

   • Top conferences