Discovering Laws of Physics via ML and Logic


Physics is fundamentally about identifying patterns in data and codifying them in mathematical language. A machine learning approach to achieve the same goal is called symbolic regression. While symbolic regression has been around for a long time, there has been some recent efforts to applying it to data to identify patterns and convert it into mathematical expressions. We propose a variation of symbolic regression, wherein, we combine a neural network-based symbolic regression tool with a mathematical reasoning solver in a corrective feedback look (a method we refer to as Logic Guided Machine Learning) and apply it to data to learn mathematical functions and physics equations. We have also done some work on pure symbolic regression (based on neural networks) to learn several symmetries and conserved quantities in physics. See publications below:


Currently, we have three projects in this space:

  1. Discovering Laws of Physics via Interpretable Siamese Neural Networks
  2. Logic Guided Machine Learning(LGML)
  3. Logic Guided Genetic Algorithms (LGGA)