Description:

Machine learning is becoming increasingly popular in science, but can these algorithms truly "understand" the scientific problems they are attempting to solve? Theory and experiment have traditionally dominated scientific research: one person constructs a clearly stated theory, which is then continuously improved upon using experimental evidence, which is then analysed to generate new predictions. But as machine learning technology has advanced quickly and the volume of scientific data has drastically increased, data-driven approaches have gained popularity. In that case, a theory already in existence is not necessary; instead, a machine learning algorithm can be utilised to examine a scientific issue using only data. The problem is, using a purely data-driven approach like this can have significant downsides. Researchers are now looking for ways to include this type of prior scientific knowledge into machine learning workflows, in the blossoming field of scientific machine learning (SciML).

I will describe physics-informed neural networks in this session, which are becoming an effective technique to apply physical concepts to machine learning. Since this a fairly new topic, research and experimentation is continuously being done on it. The goal of my talk is to invoke curiosity and awareness of such techniques among the young conference attendees, especially in India as an alternative research goal compared to the "traditional" usages of Machine Learning.

Here is a draft outline of what I will cover :

I. Introduction A. Increasing popularity of machine learning in scientific research B. Limitations of purely data-driven approaches C. Introducing the concept of physics-informed neural networks (PINNs) D. Significance of incorporating prior scientific knowledge into machine learning

II. The Need for Physics-Informed Neural Networks A. Traditional dominance of theory and experiment in scientific research B. Challenges of handling large volumes of scientific data C. Potential downsides of purely data-driven approaches

III. Understanding Physics-Informed Neural Networks (PINNs) A. Definition and core principles of PINNs B. Leveraging physical concepts to enhance machine learning models C. How PINNs bridge the gap between science and machine learning D. Overview of ongoing research and experimentation in PINNs

IV. Benefits and Applications of Physics-Informed Neural Networks A. Advantages of incorporating prior scientific knowledge B. Enhancing interpretability and explainability of machine learning models C. Potential applications in various scientific domains D. Comparing PINNs to traditional machine learning approaches

V. Conclusion A. Recap of the key points discussed B. Encouraging further research and collaboration in the field of scientific machine learning

Prerequisites:

Basic concepts of how Machine Learning and Neural Networks work in particular. Basic Physics and math concepts like differential equations, simple harmonic oscillators.

Content URLs:

Session Github repo

Research Papers : Reference 1 Reference 2

Packages I will talk about : RadonPy PINN Neural PDE

Speaker Info:

Hi I'm Vibhansh, a MLOps Engineer from the AI Team at Polymerize.io in Singapore. We help the Polymer and materials industry to accelerate their R&D using our tools. We are heavily involved in research on both the Machine Learning and Chemical sciences front.

I am also an open source contributor at PyTextRank, one of SpaCy library's pipeline extensions.

Merging the worlds of science and technology is my true passion. I have been a PyCon attendee back in 2018 when I was still a student and that is where my love for the language began. I am originally from Jammu & Kashmir. I enjoy travelling, reading and meeting new people.

Speaker Links:

Personal Profiles 1. GitHub 2. LinkedIn