Sitemap
Srashti Goyal

Srashti Goyal

Multi-Scale Problems

2 min readDec 18, 2019

The multiscale problems in physics are enormous. I must say that apart from the fundamental/basic problems in the theory and detection, what remain is mainly the multiphysics problems. What approximations to make, how much of the micro details are important while studying large scale systems, and vice versa that how the large scale flows are effecting the properties of a local system, is all the matter of multiscale physics. Even when I think of making decisions keeping in mind to impact the world, I get lost in the mechanisms. Is is very hard to compute what will happen, will it be good, is it worthy enough, am I missing on something.. etc..! And it leads to overthinking too sometimes and finally leading to the thought “Just go with the flow!”. This is what physicists are also frustrated with want it comes to determine the rainfall in a particular area as compared to the average temperature of a district. It is hard. The correlations matter. And they are not really hard to understand but hard to account for, while making predictions. There are endless no. of feedback loops. The increasing level of complexity as you dive deeper and deeper. The unknown strength of coupling constants. Or may be even uncertainty that are they really constant. Not just that, the validation of the models is also a problem. We don’t of the measurements of each and everything. We have measurements from different sets of experiments focused on different processes in a fine-tuned environments. At the very general level we will have measurements mostly at the scales that matter to us, for the predictions. The end take on it is that it is a computation problem. It is a measurement problem to. What kind approaches might be helpful then? Traditional methods include:

  1. HMM: Heterogenous multiscale methods
  2. AMR: Adaptive Mesh Resolution
  3. Hydrodynamic Limit
  4. Renormalisation group.

I would like to read up more for these methods, may be from Weinan’s book:
https://web.math.princeton.edu/~weinan/papers/weinan_book.pdf

Get Srashti Goyal’s stories in your inbox

Join Medium for free to get updates from this writer.

Finally, the exciting part is here. How can Machine learning help us in overcome or tackle this problem? The hybrid ML models seems promising when it comes to predict things in an experiment which we know should also validate the physical model that describes the processes. But due to external noise and other environmental factors it is just now exactly as expected. This is one of the ways in which I find physics and ML can be combined and in the end lead to something more powerful. One of the main issues is that we don’t know the initial conditions, or we don’t have the measurements everywhere. I have a hunch that ML can help in overcoming the computational problem, maybe not the measurement problem, but that can be generated using physical model numerical simulations.

For example: Can we have a ML model that takes input from NSE simulations at different resolutions for various Reynolds no. and then in the end tells the dynamics at the higher/intermediate Reynolds numbers.

--

--