If we switch to GUIs, the experimentalists might realise how easy we have it!

If I've misunderstood anything, I do apologize. Please correct me in the thread. Overall I think the paper's going to do very well and I look forward to reading the final print! 12/12

reduce the number of force evaluations required per optimisation, then their method is likely to be overall cheaper than contemporary CSP methods and I would expect to see the model as a strong contender in the next CSP blind test. 11/12

There is one final point in which I think that they're probably *under*selling their work. The accuracy of their generated structures is likely to improve if optimised with DFT. If their model can reduce the number of structures that need to be optimised, and 10/12

I'd really like to see how the model compares to the blind test participants using the 7th blind test's criteria for a match. I would also be really interested to see how Sol_c changes with the number of crystals generated and also the number of symmetry-independent molecules. 9/12

They do summarise the performance of each group for every target in their SI but they haven't reported the performance of OXtal on each target so we can't compare. 8/12

If they're averaging over each group's performance, that's flawed in the other way because the best performing groups/methods are impeded my the lower performing groups/methods. 7/12

If they're counting a successful prediction by any group as a success for "DFT-based submissions" then they're probably being overly generous to DFT because they're sampling across an ensemble of different DFT approaches. 6/12

They compare their model initially to other generative models on CSD structures but later compare their performance against a "aggregate of reported expensive DFT-based submissions" for previous CSP blind tests. I'm not entirely sure how they're aggregated though. 5/12

In the preprint, the criteria for "approximately solved" (Sol_c) was an RMSD of less than 2 Å for a cluster of 15 molecules (a significantly looser tolerance). They're not using the word "match" here, but it really feels implied that that's what they have. 4/12

The authors use the @ccdc.cam.ac.uk's COMPACK metric to overlay crystals and minimise the RMSD between clusters of N molecules. COMPACK is also used in the CSP blind tests but in the latest blind test, the criteria for a match was an RMSD of less than 1 Å for a cluster of 30 molecules. 3/12

Unfortunately, the data as presented makes me feel that the claims made by the authors are somewhat exaggerated. I'm particularly concerned with the metrics which are described as "standard CSP metrics". 2/12

New addition for the Day group office?

Choose comp-chem and you won't even have to go into the office!

Strange way to ask what sector of science people work in.