Programming

The reason they don't do multi people and multi year coding projects has nothing to do with repeatability of the experiments, most science coding is done through simple-ish code that uses existing libraries, doesn't code them. That code is usually stored in notebooks (jupyter, zeppelin) or simple scripts.

For science code, it usually falls in the realm of data analysis, and as a data engineer, let me tell you that the analysis part of the job is usually very ad-hoc modifications of the script and live coding though notebooks and such.

The part where whatever conclusion of the research is then transformed into a functioning application, taking care of naming conventions, the architecture of the system where the input data, the transformations, the postprocessing and such is done, is usually done by another team of dedicated data engineers or software developers.

I guess that it would be helpful for the analysis part to have standardized templates for data extraction and such, but usually the tools used in the research portion of the process and the implementation portion are completely different (python with tensorflow vs C++ with openvino or whatever cloud based) so it's not really fair to load the architecture design since the beginning.

You know how changing requirements is the bane of real™ production grade™ software?

In science requirements change all the time. You write some 50-100 lines to plot your results. You realize that the effect is not visible, so you go back to the lab, change 5 variables and run the test again. Some quick code changes and you see the effect. Perfect. Now you do the measurement as a function of temperature. You adapt your script, you indent the data processing code to turn your list of files into a list of characteristics parameters and adapt the plotting. You run the experiment for the three samples you have prepared and compare their plots. Some more experiments tuning and corresponding script tuning is required. You take the characteristic parameters that your code (grew to 200 lines now, but whatever) calculated and write a new script to take that array and plot it nicely.

Now someone wants to repeat the experiment 4 years later. The measurement equipment changed and the data format is slightly different. It's impossible to document the exact state of the hardware in your code anyway. They are actually interested in a different effect and want to plot that as well, but they need the effect/characteristics parameters that are shown by your code as a sanity check. They need to rewrite 125 of the 200 lines.

There never is a finished product that is worth maintaining long term. Everyone using the script has to understand the domain precisely anyway. Is it worth it to reuse the old code when you need to rewrite more than half of it anyway?

Don't get me wrong, code reuse does happen. But it's much more "oh, I wrote that three months ago somewhere else" ctrl-c ctrl-v. "Ok, now I need to change five lines in this function to adapt to the new thing I'm trying to do." It makes absolutely no sense to write that function in a abstracted way. Every time you use it the requirements changed and the abstraction is no longer valid anyway.

Yes, those prototypes are the goal here.