In Praise of Visualization

A little investment into building a GUI for live visualizations and control provides huge improvements in iteration speed.

I recently came across the excellent video Coding Adventure: Simulating Fluids wonderfully explained by Sebastian Lague. The video walks through the steps to build a simple smoothed particle hydrodynamics simulation from nothing, in C#.

It brings back some sweet memories to my days as a research student doing discrete particle model simulations using MercuryDPM and there are many practical lessons that I wish I had known when I first started. The one that stuck out the most was actually one that Lague glossed over: the power of visualisation and the massive returns that even a simple user interface gives.

The most dull and most frustrating aspect of my computational work was undoubtedly the parameter studies to tune parameters such as time step or particle stiffness. Poorly-chosen values do not immediately cause problems; your simulation will crank along seemingly fine for several minutes before blowing up, or having nans everywhere).

I didn't have a good visualization system at the time (MercuryDPM does provide options for loading results into visualization software, but it isn't live) and this hugely slowed down my iteration cycles. Neither did I have a GUI for modifying parameters. So, my iterations were rate-limited by my ability to load files into gnuplot or investigate them manually.

I'm now mostly not working on simulations, instead working almost full-time in data engineering. The lesson around visualization is the same, though: even a simple GUI can give a lot of insight and control over what is otherwise a black box system, allowing faster iterations.

Visualization systems, like other pieces of "infrastructure" such as testing frameworks, do not per se produce results, and so are often overlooked or disincentivised in environments that prioritize rolling out a product or result (proverbially, "moving fast and breaking things"), especially research environments. I have been guilty of this myself, but Sebastian's video clearly demonstrates that introducing even a small amount of interactivity can be very powerful when guiding one towards building reliable software and choosing sensible parameters. I wish I had been a stronger advocate of this as a junior researcher or engineer.

Perhaps building a GUI is easiest in a language like C# or JavaScript. (I have no experience with the former.) I have yet to find a GUI framework in Python that isn't either overly restrictive or horribly low-level. Jupyter notebooks offer some hope in this area, but I might look into PyGame.