Mine does the "thinking" thing and then answers with just a (correct) number.

Same for me. Now that there are pretty good website builders out there I will switch to a simple markdown based system next year and host for free on cloudflare pages or similar. Involves some coding but it's very limited.

This is still an early version with limited functionality — mostly built around what we need in our own research.

Right now, the focus is on ORCA 6 compound output files, especially those using compound scripts. Support for other formats and features will grow with feedback and community input.

We are currently finishing up that work, I might contact you directly if stuff like that is of interest.

Thanks for your insights! Of course you are always limited by the tools. We have a quite simple solution to the cis/trans problem in structure Generation that I think could be integrated into agentic workflows like this pretty nicely.

I guess I have to stop by and try some of our transition states. ;-) Would be interesting to implement it in our workflows.

wrong structures would be used, similar with the trans-cyclooctene and cycloheptynes.

More complex problems involve finding transition states, e.g. between the trans-cycloheptene and 1,2,4,5-tetrazine, or mechanistic studies with very shallow barriers.

Hi Yunheng,
Thanks for reaching out. A seemingly simple example we had troubles with automated systems before would be "perform a geometry optimization of trans-cycloheptene C1C/C=C/CCC1 using GFN2-xTB". Default obabel and rdkit struggle with generating structures like this, unless there are checks

That sounds interesting. We do some more complex calculation workflows/have some weird molecules and I wonder if it can deal with that.

Each country has their own research grants too... And in the US grants have huge overheads and PIs have to pay tuition for their students. In the EU universities are mostly funded directly and research grants only pay for research expenses and salaries with little or even no overhead.

Are people actually using ML force fields if physics based methods are still viable? We use the ML once the systems get too big/MD too long or GFN2-xTB cannot handle the systems at all (e.g., Actinides) and from what I can tell that's usually also the case if people use ACE/MACE.