The paper touches on how those impact on large datasets for later training. I am glad the effort was worth it. Honestly both were a lot of work and didn't come without sweat and tears 😅.

OBS.: DEFGRID3 is definitely recommended for benchmark data, that's why we made it!

3/3

We have first have developed the new ML-optimized grids for numerical integration in ORCA 5, which have smaller error even with less points.

Then derivied and implemented a fully translation-invariant COSX gradient for ORCA 6, which reduces numerical noise.

2/3

😂

PS3: The original post has only limited context (which is OK), so my comment here is not about that one in particular, just in general since so much is happening, and you did ask for it 😆.

PS2: The accuracy I refer to is for MLIP. New functionals like the Skala do improve over wB97M-V - but that's still regular DFT, just a better functional!

They have accelerated what people have been doing by hand for years and it's a perfectly reasonable approach to finding the XC func., IMHO.

PS.: Not even the authors made such big claims, they did a really nice job there and reported as it is. I hope we can test all this and put it to use for the benefit of us all.

Yes, I don't even know why we are discussing this in 2025, as if any of these models they could extrapolate far from the training. This is good old magical thinking, certainly not mathematical one.

5. Overarching claims, IMHO, actually do not help the field, because people might loose trust once it fails. And it will fail, as everything does, so there's no need for that.

We don't need to revolutionize the world every other week. Continuous progress is fine too 😀.

4. It does not mean they are no useful, it is actually amazing that this was achieved! These method will probably become part of our toolkit to solve problems, and they can help a lot. Specially for things like geometry optimization and MD, could be really something.

3. As we know from other models, the gains get smaller with the size of the system. So even to get to the "next level" wB97M-V quality, one probably needs orders of magnitude more parameters and training size, which would be quite challenging. CC seems far off in the horizon.

2. Even if it could perfectly extrapolate to the entire chemical space, the current quality is of good non-hybrid DFT like r2SCAN-3c. It does not "solve the problem", as we know non-hybrid DFT does not unfortunately.

I don't know exactly what Prof. Reiher meant with this to be honest. But in general, I keep my take on it:

1. All current AI/ML are interpolation methods. Which means that, by construction, they are bound to the training space, or "points between points" in the dataset, maybe a bit further.

Thanks, we are trying :)

In the end it is quite trivial. But I can tell, from a lot of experience, that 33 kcal/mol from such small changes is definitely not the usual.

It really makes me think how "high in energy" might be some drug poses be inside a protein pocket and how much we miss if we are not looking into them.

That's also a good idea. We need to use CovaLED to break through single bonds, but certainly possible.

In this case we just wanted to see if there was one specific fragmengt with a clearly defined spike in energy, but there wasn't.

Take a look at the structures, it's quite interesting