Dear Henrique, Yes, my previous representation looks like a Z-matrix format (BTW in scikit-learn you will need to have the same number of columns at every line, so you will need to fill somehow the first line). However, I will take this email as the opportunity to stress the fact that you do not have to stick to a specific file format, but the the features/columns/(2nd index of a 2D matrix) they have to represent the properties/parameters that they will directly affect changes in what you are trying to predict. In fact in your columns you can even add more features compared to the priviosly cited and probably you will describe the system in a better way. What I can think right now, just for the seak of a better explanation, you can use: bond lenght, atoms type, number of unpaired electrons, total number of electrons, diedral angle of the two atoms, number of atoms between the pair (e.g. if you have Mn--O--Mn there is an oxygen between the two Mn that you might want to look at the coupling) and so on and so forth. The number of parameters you will have to use will solely depend on your system and what you need to describe, but it will not affect in any case scikit-learn routines. Basically every 2D matrix of number will work on scikil-learn, but in order to make these number to have physical meaning that will depend on what the number represent. Let me know if it make more sense Sincerely Tommaso On Mar 28, 2017 12:51 PM, "Henrique C. S. Junior" <henriquecsj@gmail.com> wrote: @Tommaso, this is something like Internal Coordinates[1], right? @Bill, thanks for the hint, I'll definitely take a look at this. [1] - https://en.wikipedia.org/wiki/Z-matrix_(chemistry) On Tue, Mar 28, 2017 at 2:12 AM, Bill Ross <ross@cgl.ucsf.edu> wrote:

Image processing deals with xy coordinates by (as I understand) training with multiple permutations of the raw data, in the form of translations and rotations in the 2d space. If training with 3d data, there would be that much more translating and rotating to do, in order to divorce the learning from the incidentals.

Bill

On 3/27/17 4:35 PM, Tommaso Costanzo wrote:

Dear Henrique, I am sorry for the poor email I wrote before. What I was saying is simply the fact that if you are trying to use the coordinates as "features" from an .xyz file then by machine learning you will learn at wich coordinate certain atoms will occur so you can only make prediction on the coordinate. However, if I correctly understood, the "features" representing the coupling J are distance, angle, and electron number. Definitely this properties can be derived from the XYZ file format from simple geometric calculations and the number of electrons will depend from the type of atom. So, what I was trying to say is that instead of using the XYZ file as input for scikit-learn, I was suggesting to do the calculation of angle, distances, electrons' number in advance (with other software(s) or directly in python) and use the new calculated matrix as input for scikit-learn. In this case the machine will learn how J(AB) varies as a function of angle, distance, number of electrons. For example

distance angle n el. 1 90 1 1 90 1 2 90 1 .... ... ...

If you are using a supervised learning you will have to add a 4th column ( in reality a separate column vector) with your J(AB) on which you can train your model and then predict the unknown samples

For example distance angle n el. J(AB) 1 90 1 1 1 90 1 1 2 90 1 0.5 .... ... ... ...

Now if you train the model on the second matrix, and then you try to predict the first one you should expect a results like:

1 1 0.5

Of course in this case the "features" are perfectly equal, hence the example is completely unrealistic. However, I hope that it will help to understand what I was explaining in the previous email. If you want you can directly contact me at this email, and I hope that you got additional hints from Robert, that he seems to be even more knowledgeable than me.

Sincerely Tommaso

2017-03-27 18:44 GMT-04:00 Henrique C. S. Junior <henriquecsj@gmail.com>:

...

Dear Tommaso, thank you for your kind reply. I know I have a lot to study before actually starting any code and that's why any suggestion is so valuable. So, you're suggesting that a simplification of the system using only the paramagnetic centers can be a good approach? (I'm not sure if I understood it correctly). My main idea was, at first, try to represent the systems as realistically as possible (using coordinates). I know that the software will not know what a bond is or what an intermolecular interaction is but, let's say, after including 1000s of examples in the training, I was expecting that (as an example) finding a C 0.000 and an H at 1.000 should start to "make sense" because it leads to an experimental trend. And I totally agree that my way to represent the system is not the better.

Thank you so much for all the help.

On Mon, Mar 27, 2017 at 4:15 PM, Tommaso Costanzo < tommaso.costanzo01@gmail.com> wrote:

...

Dear Henrique,

I agree with Robert on the use of a supervised algorithm and I would also suggest you to try a semisupervised one if you have trouble in labeling your data.

Moreover, as a chemist I think that the input you are thinking to use is not the in the best form for machine learning because you are trying to predict coupling J(AB) but in the future space you have only coordinates (XYZ). What I suggest is to generate the pair of atoms externally and then use a matrix of the form (Mx3), where M are the pairs of atoms you want to predict your J and 3 are the features of the two atoms (distance, angle, unpaired electrons). For a supervised approach you will need a training set where the J is know so your training data will be of the form Mx4 and the fourth feature will be the J you know.

Hope that this is clear, if not I will be happy to help more

Sincerely

Tommaso

2017-03-27 13:46 GMT-04:00 Henrique C. S. Junior <henriquecsj@gmail.com> :

...

Dear Robert, thank you. Yes, I'd like to talk about some specifics on the project. Thank you again.

On Mon, Mar 27, 2017 at 2:25 PM, Robert Slater <rdslater@gmail.com> wrote:

...

You definitely can use some of the tools in sci-kit learn for supervised machine learning. The real trick will be how well your training system is representative of your future predictions. All of the various regression algorithms would be of some value and you make even consider an ensemble to help generalize. There will be some important questions to answer--what kind of loss function do you want to look at? I assumed regression (continuous response) but it could also classify--paramagnetic, diamagnetic, ferromagnetic, etc...

Another task to think about might be dimension reduction. There is no guarantee you will get fantastic results--every problem is unique and much will depend on exactly what you want out of the solution--it may be that we get '10%' accuracy at best--for some systems that is quite good, others it is horrible.

If you'd like to talk specifics, feel free to contact me at this email. I have a background in magnetism (PhD in magnetic multilayers--i was physics, but as you are probably aware chemisty and physics blend in this area) and have a fairly good knowledge of sci-kit learn and machine learning.

On Mon, Mar 27, 2017 at 10:50 AM, Henrique C. S. Junior < henriquecsj@gmail.com> wrote:

...

I'm a chemist with some rudimentary programming skills (getting started with python) and in the middle of the year I'll be starting a Ph.D. project that uses computers to describe magnetism in molecular systems.

Most of the time I get my results after several simulations and experiments, so, I know that one of the hardest tasks in molecular magnetism is to predict the nature of magnetic interactions. That's why I'll try to tackle this problem with Machine Learning (because such interactions are dependent, basically, of distances, angles and number of unpaired electrons). The idea is to feed the computer with a large training set (with number of unpaired electrons, XYZ coordinates of each molecule and experimental magnetic couplings) and see if it can predict the magnetic couplings (J(AB)) of new systems: (see example in the attached image)

Can Scikit-Learn handle the task, knowing that the matrix used to represent atomic coordinates will probably have a different number of atoms (because some molecules have more atoms than others)? Or is this a job better suited for another software/approach? ​

-- *Henrique C. S. Junior* Industrial Chemist - UFRRJ M. Sc. Inorganic Chemistry - UFRRJ Data Processing Center - PMP Visite o Mundo Químico <http://mundoquimico.com.br>

_______________________________________________ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn

_______________________________________________ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn

-- *Henrique C. S. Junior* Industrial Chemist - UFRRJ M. Sc. Inorganic Chemistry - UFRRJ Data Processing Center - PMP Visite o Mundo Químico <http://mundoquimico.com.br>

_______________________________________________ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn

-- Please do NOT send Microsoft Office Attachments: http://www.gnu.org/philosophy/no-word-attachments.html

_______________________________________________ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn

-- *Henrique C. S. Junior* Industrial Chemist - UFRRJ M. Sc. Inorganic Chemistry - UFRRJ Data Processing Center - PMP Visite o Mundo Químico <http://mundoquimico.com.br>

_______________________________________________ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn

-- Please do NOT send Microsoft Office Attachments: http://www.gnu.org/philosophy/no-word-attachments.html

_______________________________________________ scikit-learn mailing listscikit-learn@python.orghttps://mail.python.org/mailman/listinfo/scikit-learn

_______________________________________________ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn

-- *Henrique C. S. Junior* Industrial Chemist - UFRRJ M. Sc. Inorganic Chemistry - UFRRJ Data Processing Center - PMP Visite o Mundo Químico <http://mundoquimico.com.br> _______________________________________________ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn