\"Using state-of-the-art methods from quantum mechanics, it is possible to accurately calculate scalar coupling constants given only a 3D molecular structure as input.\"
\n\nNow the challenge is over, I am still interested in how the J coupling constant are generated.\nIt would be nice if the method can be disclosed (e.g. which quantum mechanics software, what atom basis, other parameters, etc)
\n\nI find it a little strange that the bond values (e.g. 1JHC, etc) given here are not what i would get in the typical textbook. I suspect the some parameters are different (e.g. basis, temperature )?
\n\nThanks!
", "messages": [ { "id": "610851", "postDate": "08/29/2019 02:02:04", "content": "\"Using state-of-the-art methods from quantum mechanics, it is possible to accurately calculate scalar coupling constants given only a 3D molecular structure as input.\"
\n\nNow the challenge is over, I am still interested in how the J coupling constant are generated.\nIt would be nice if the method can be disclosed (e.g. which quantum mechanics software, what atom basis, other parameters, etc)
\n\nI find it a little strange that the bond values (e.g. 1JHC, etc) given here are not what i would get in the typical textbook. I suspect the some parameters are different (e.g. basis, temperature )?
\n\nThanks!
", "rawMarkdown": "\"Using state-of-the-art methods from quantum mechanics, it is possible to accurately calculate scalar coupling constants given only a 3D molecular structure as input.\"\n\nNow the challenge is over, I am still interested in how the J coupling constant are generated.\nIt would be nice if the method can be disclosed (e.g. which quantum mechanics software, what atom basis, other parameters, etc)\n\nI find it a little strange that the bond values (e.g. 1JHC, etc) given here are not what i would get in the typical textbook. I suspect the some parameters are different (e.g. basis, temperature )?\n\nThanks!", "votes": null }, { "id": "611676", "postDate": "08/29/2019 11:30:48", "content": "The scalar coupling constants was computed in vacuum using DFT with b3lyp/6-31g(2df,p) in Gaussian09E. This is the same methodology used to create the QM9 dataset and this was chosen such that this dataset could compliment the QM9 dataset for benchmarking future ML models in the community. \nThe discrepancy between the computed values and experiment is mainly due to three reasons.\n1) It is well known that DFT with a small basis set have a systematic offset to experimental data, since the smaller basis set doesn't adequately describe the electron distribution and DFT has some deficiencies compared to more accurate but slower methods.\n2) These values were computed in vacuum, rather than in solvent. Solvent would have a very big effect especially on the more polar atoms, such as in 1JHN.\n3) Experimental observables is an average of several molecules that move around in solution. The values in the competition is computed for a single minimized structure.
\n\nThis is all something that you can account for in some way in practical use, but we hypothesized that none of these should greatly impacton what types of algorithms perform well.
", "rawMarkdown": "The scalar coupling constants was computed in vacuum using DFT with b3lyp/6-31g(2df,p) in Gaussian09E. This is the same methodology used to create the QM9 dataset and this was chosen such that this dataset could compliment the QM9 dataset for benchmarking future ML models in the community. \nThe discrepancy between the computed values and experiment is mainly due to three reasons.\n1) It is well known that DFT with a small basis set have a systematic offset to experimental data, since the smaller basis set doesn't adequately describe the electron distribution and DFT has some deficiencies compared to more accurate but slower methods.\n2) These values were computed in vacuum, rather than in solvent. Solvent would have a very big effect especially on the more polar atoms, such as in 1JHN.\n3) Experimental observables is an average of several molecules that move around in solution. The values in the competition is computed for a single minimized structure.\n\nThis is all something that you can account for in some way in practical use, but we hypothesized that none of these should greatly impacton what types of algorithms perform well.", "votes": null }, { "id": "611964", "postDate": "08/29/2019 14:30:44", "content": "Can I ask what kind of hardware it took to calculate all of the constants? And how long it took?
", "rawMarkdown": "Can I ask what kind of hardware it took to calculate all of the constants? And how long it took?", "votes": null }, { "id": "612063", "postDate": "08/29/2019 15:28:52", "content": "Very rough estimate since the calculations were split over different super computers (all Xeons less than 5 years old though), but somewhere between 200-300K core hours.
", "rawMarkdown": "Very rough estimate since the calculations were split over different super computers (all Xeons less than 5 years old though), but somewhere between 200-300K core hours.", "votes": null }, { "id": "612106", "postDate": "08/29/2019 15:53:49", "content": "I have two questions:\n- Did you guys use Kohn-Sham model?\n- After the basis sets are plugged into the model of choice, you have to solve a dense linear system right? What solver you guys use?\nThanks.
", "rawMarkdown": "I have two questions:\n- Did you guys use Kohn-Sham model?\n- After the basis sets are plugged into the model of choice, you have to solve a dense linear system right? What solver you guys use?\nThanks.", "votes": null }, { "id": "612128", "postDate": "08/29/2019 15:59:24", "content": "I didn't write the QM code myself, so I used whatever solver Gaussian uses. But yes, DFT uses the Kohn-Sham equations.
", "rawMarkdown": "I didn't write the QM code myself, so I used whatever solver Gaussian uses. But yes, DFT uses the Kohn-Sham equations.", "votes": null }, { "id": "612178", "postDate": "08/29/2019 16:30:22", "content": "@larsbratholm Wow! That's a lot of core hours. Thanks for the information!
", "rawMarkdown": "larsbratholm Wow! That's a lot of core hours. Thanks for the information!", "votes": null } ], "comments": [ { "id": 611676, "author_name": "larsbratholm", "author_url": "", "post_date": "08/29/2019 11:30:48", "content": "The scalar coupling constants was computed in vacuum using DFT with b3lyp/6-31g(2df,p) in Gaussian09E. This is the same methodology used to create the QM9 dataset and this was chosen such that this dataset could compliment the QM9 dataset for benchmarking future ML models in the community. \nThe discrepancy between the computed values and experiment is mainly due to three reasons.\n1) It is well known that DFT with a small basis set have a systematic offset to experimental data, since the smaller basis set doesn't adequately describe the electron distribution and DFT has some deficiencies compared to more accurate but slower methods.\n2) These values were computed in vacuum, rather than in solvent. Solvent would have a very big effect especially on the more polar atoms, such as in 1JHN.\n3) Experimental observables is an average of several molecules that move around in solution. The values in the competition is computed for a single minimized structure.
\n\nThis is all something that you can account for in some way in practical use, but we hypothesized that none of these should greatly impacton what types of algorithms perform well.
", "votes": null, "replies": [ { "id": 611964, "author_name": "abendorf", "author_url": "", "post_date": "08/29/2019 14:30:44", "content": "Can I ask what kind of hardware it took to calculate all of the constants? And how long it took?
", "votes": null, "replies": [] }, { "id": 612063, "author_name": "larsbratholm", "author_url": "", "post_date": "08/29/2019 15:28:52", "content": "Very rough estimate since the calculations were split over different super computers (all Xeons less than 5 years old though), but somewhere between 200-300K core hours.
", "votes": null, "replies": [] }, { "id": 612106, "author_name": "scaomath", "author_url": "", "post_date": "08/29/2019 15:53:49", "content": "I have two questions:\n- Did you guys use Kohn-Sham model?\n- After the basis sets are plugged into the model of choice, you have to solve a dense linear system right? What solver you guys use?\nThanks.
", "votes": null, "replies": [] }, { "id": 612128, "author_name": "larsbratholm", "author_url": "", "post_date": "08/29/2019 15:59:24", "content": "I didn't write the QM code myself, so I used whatever solver Gaussian uses. But yes, DFT uses the Kohn-Sham equations.
", "votes": null, "replies": [] }, { "id": 612178, "author_name": "abendorf", "author_url": "", "post_date": "08/29/2019 16:30:22", "content": "@larsbratholm Wow! That's a lot of core hours. Thanks for the information!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "610851": "\"Using state-of-the-art methods from quantum mechanics, it is possible to accurately calculate scalar coupling constants given only a 3D molecular structure as input.\"\n\nNow the challenge is over, I am still interested in how the J coupling constant are generated.\nIt would be nice if the method can be disclosed (e.g. which quantum mechanics software, what atom basis, other parameters, etc)\n\nI find it a little strange that the bond values (e.g. 1JHC, etc) given here are not what i would get in the typical textbook. I suspect the some parameters are different (e.g. basis, temperature )?\n\nThanks!", "611676": "The scalar coupling constants was computed in vacuum using DFT with b3lyp/6-31g(2df,p) in Gaussian09E. This is the same methodology used to create the QM9 dataset and this was chosen such that this dataset could compliment the QM9 dataset for benchmarking future ML models in the community. \nThe discrepancy between the computed values and experiment is mainly due to three reasons.\n1) It is well known that DFT with a small basis set have a systematic offset to experimental data, since the smaller basis set doesn't adequately describe the electron distribution and DFT has some deficiencies compared to more accurate but slower methods.\n2) These values were computed in vacuum, rather than in solvent. Solvent would have a very big effect especially on the more polar atoms, such as in 1JHN.\n3) Experimental observables is an average of several molecules that move around in solution. The values in the competition is computed for a single minimized structure.\n\nThis is all something that you can account for in some way in practical use, but we hypothesized that none of these should greatly impacton what types of algorithms perform well.", "611964": "Can I ask what kind of hardware it took to calculate all of the constants? And how long it took?", "612063": "Very rough estimate since the calculations were split over different super computers (all Xeons less than 5 years old though), but somewhere between 200-300K core hours.", "612106": "I have two questions:\n- Did you guys use Kohn-Sham model?\n- After the basis sets are plugged into the model of choice, you have to solve a dense linear system right? What solver you guys use?\nThanks.", "612128": "I didn't write the QM code myself, so I used whatever solver Gaussian uses. But yes, DFT uses the Kohn-Sham equations.", "612178": "larsbratholm Wow! That's a lot of core hours. Thanks for the information!" }, "source": "meta" }