Comment by HiPHInch

Comment by HiPHInch 3 days ago

11 replies

View on Hacker News

I took some effort to change my research interest from computer vision to DFT calculation in quantum chemistry.

Honestly, I'm kind of frustrated now, too many work is close-source in this area. The research paper will tell you everything except how to reproduce this work in minimal effort, it's like they are hiding something.

They also using a `Origin` to plot and MS Word to write paper, which is also non-free licensed, and made them harder to collaborate and reproduce.

BeetleB 3 days ago

> The research paper will tell you everything except how to reproduce this work in minimal effort, it's like they are hiding something.

They are. I used to work in an adjacent field. Everyone was open about doing it - they're competing with others for grants, and worry that if they reveal the secret sauce, others will move faster than they can.

You can say you performed a DFT calculation to get the result, but anyone who's studied these types of simulations/calculations knows that it's highly nontrivial to implement, with lots of coding and numerical tricks involved. So it's extremely hard to reproduce if you don't have detailed access to the algorithms.

Reply View 2 replies
  • malux85 2 days ago

    Not only that, but DFT itself has many many different forms. There's DFT that is O(n)^3 and there's DFT that's O(n)^7 in time complexity, the wild variations are due to the different approximations (i.e. algorithm and parameters).

    Saying "I used DFT" is like saying "I used a computer", its nowhere near enough info to reproduce the work

    Reply View 1 reply
    • HiPHInch 2 days ago

      You are right. But I think nearly a half of DFT calcs are done with VASP.

      Particularly, if one using ASE or other higher level wrapper of calculators(like quacc), he can share all the params in just one python script.

      If not, just share the INCARs and POSCARs using a github link or whatsoever.

      Reply View 0 replies
wholinator2 2 days ago

Very true that they're hiding things. I actually wrote some code (that strung together other people's code) to complete a simulation pipeline for non adiabatic molecular dynamics. I was tasked with writing documentation to teach the group but was instructed to not release it anywhere publicly because other groups would simply take the method and move faster since they had more money and compute.

Reply View 0 replies
gaugefield 3 days ago

This issue also bugged me for a while. It is more of cultural issue, and older the research group is, the less likely it is for research software to be open, in my experience.

In the area of deep learning based simulations, one good example of an open software is netket. The researcher their is pretty active in terms of github/gitlab/huggingface ecosystem.

Reply View 0 replies
qwezxcrty 3 days ago

I miss OriginPro in my undergrad when we had campus licenses for, before moving to matplotlib for data visualization. matplotlib is simply too disappointing for making publication quality figures. The most recently encountered problem is how to plot with a broken x-axis, which is one of the most basic need in physical science but requires a non-trivial amount of hacking to get with matplotlib.

Open source tool or not, I don't care at all as I get the science right. I have already enough frustration dealing with my samples, so I simply want the least frustration from the software I use to plot.

Reply View 4 replies
  • prennert 3 days ago

    Matplotlib is a bit painful. Often seaborn will work quicker, especially when using Pandas dataframes with proper column names and seaborn compatible layout.

    Its annoying that you cannot create a broken axis out-of-the box, but I am sure you can wrap this to make your own convenience function: https://matplotlib.org/stable/gallery/subplots_axes_and_figu...

    Reply View 1 reply
    • qwezxcrty 2 days ago

      That link was what I referred to after Googling, but in my case I need the width of the left part and the right part to be different, which requires setting width_ratios in the subplots and adjusting the slope of the hacky lines used to draw the broken axis symbol. seaborn also would not help in this exact case.

      There is a package by some nice guy: https://github.com/bendichter/brokenaxes just to do the broken axis. But not being built-in in Anaconda is already an annoyance, and in my case it generates a figure with a ugly x-label.

      I ended up letting ChatGPT generate the code for me with the two required hacks. I simply need the figure in the minimal amount of time and with the least mental bandwidth, so I can focus on the science and catch the conference deadline. Origin is a very "over-engineered" piece of software, but hey getting a broken axis is so simple (https://www.originlab.com/doc/Origin-Help/AxesRef-Breaks ). Sometimes the "over-engineering" is necessary to minimize users' pain.

      Reply View 0 replies
  • mvieira38 2 days ago

    Honestly, if you're doing scientific work there is no reason not to output the data somewhere and plot in R with the standard lib (insanely good for science style plotting but hard to use) or ggplot (what matplotlib wished it was)

    Reply View 0 replies
  • foven 2 days ago

    Honestly, when it comes to hacking things together with matplotlib I outsource all of my thinking to chatgpt to do the 80% of doc hunting that is honestly not worth it since everything in matplotlib is labelled inconsistently.

    Reply View 0 replies