I didn't follow. If you take a high dimensional data set and project it are you not going to find patterns? In fact, let me ask it differently. Are there any projections that don't find patterns?
Right. He mentions a few but they again seem to be things chemists were aware of. I really wanted to see the patterns chemists didn't think of instead of the algorithms recovering what was already known.
This is very far from a rediscovery. The program just rearrange the elements with some criteria in the plane.
The important part of the periodic table was that it is a regular arrangement. The elements in the same column and rows share some properties, you can somewhat guess the properties of an element looking at the properties of the nearby elements.
For example, all the elements in the first column are alkaline, they form salts with chloride like XCl. The solubility, density, ... of these salts are somewhat related. Some columns have a clear relation, some columns have more messy relations.
The relation in each row is less useful, only atomic weight, radius and electronegativity, but there may be some special cases an exceptions.
A clear grid arrangement by Mendeleev has two important properties:
* He first ordered the elements by atomic weight, but it has a few weird cases, so he decided to rearrange them. This rearrangement was later explained when the internal structure of the atoms was understood.
* Some slots in the grid were missing, so he predicted some new elements and their properties. These elements were discover later, with properties similar to the predicted.
It's not very clear how much insight these new representation provide. To make the task more clear: If these algorithm use only the data available to Mendeleev, can they discover germanium and gallium?
"The elements in the same column and rows share some properties"
While this is what most are taught, in my opinion this is not the best way to understand it and this can actually lead to mistakes.
The chemical properties of the elements are mostly determined by 3 characteristics: valence, electronegativity and size.
Valence and electronegativity are usually similar for elements in the same column and size is usually similar for elements in the same row.
This is the cause for the similarities in columns and rows. However, while the electronegativity and the size have quasi-periodic variations with the atomic number, there are various reasons that shift their similar values in adjacent rows, so that there are many cases when there are much more similarities between elements in different columns than between those in the same column. Usually for the heavy elements grouping them by column is much more useful than for the light elements, which can be more similar to elements in the same row or to elements in a diagonal position.
For example, due to their closer electronegativity and size, carbon and sulfur resemble much more to each other than to silicon or oxygen.
As another of the many examples, in the first row of the transition metals the maximum density and the change from hexagonal to cubic structure is located at cobalt-nickel, but in the next 2 rows it is shifted to ruthenium-rhodium and osmium-iridium.
There is a huge number of such examples, so that anyone who wants to understand the similarities between chemical elements must not use directly the column and row numbers but must know how the electronegativity, the atomic size and the valence vary within the periodic table.
I agree. The column bellow Boron is a zoo. (Boron probably deserve a category for itself.) The non-metal part is delimited by a diagonal (and the border is fuzzy).
Some columns are useful, I, II, VIII, VII, perhaps VI. That's at least half of the main groups. Other parts are not so clear. But it is possible to use the grid structure to make some approximations and predictions of the properties.
TLDR; To figure out property of chemical composition of one specific form there is a method called DFT which is expensive. There are 10^6 combinations for this molecule composition from elements. So they train ridge-regression model to predict it instead of DFT and find some interesting compositions.
I wonder if it is possible to find much more general molecules that might have interesting properties through ML.
8 comments
[ 2.9 ms ] story [ 31.1 ms ] threadThe important part of the periodic table was that it is a regular arrangement. The elements in the same column and rows share some properties, you can somewhat guess the properties of an element looking at the properties of the nearby elements.
For example, all the elements in the first column are alkaline, they form salts with chloride like XCl. The solubility, density, ... of these salts are somewhat related. Some columns have a clear relation, some columns have more messy relations.
The relation in each row is less useful, only atomic weight, radius and electronegativity, but there may be some special cases an exceptions.
A clear grid arrangement by Mendeleev has two important properties:
* He first ordered the elements by atomic weight, but it has a few weird cases, so he decided to rearrange them. This rearrangement was later explained when the internal structure of the atoms was understood.
* Some slots in the grid were missing, so he predicted some new elements and their properties. These elements were discover later, with properties similar to the predicted.
It's not very clear how much insight these new representation provide. To make the task more clear: If these algorithm use only the data available to Mendeleev, can they discover germanium and gallium?
While this is what most are taught, in my opinion this is not the best way to understand it and this can actually lead to mistakes.
The chemical properties of the elements are mostly determined by 3 characteristics: valence, electronegativity and size.
Valence and electronegativity are usually similar for elements in the same column and size is usually similar for elements in the same row.
This is the cause for the similarities in columns and rows. However, while the electronegativity and the size have quasi-periodic variations with the atomic number, there are various reasons that shift their similar values in adjacent rows, so that there are many cases when there are much more similarities between elements in different columns than between those in the same column. Usually for the heavy elements grouping them by column is much more useful than for the light elements, which can be more similar to elements in the same row or to elements in a diagonal position.
For example, due to their closer electronegativity and size, carbon and sulfur resemble much more to each other than to silicon or oxygen.
As another of the many examples, in the first row of the transition metals the maximum density and the change from hexagonal to cubic structure is located at cobalt-nickel, but in the next 2 rows it is shifted to ruthenium-rhodium and osmium-iridium.
There is a huge number of such examples, so that anyone who wants to understand the similarities between chemical elements must not use directly the column and row numbers but must know how the electronegativity, the atomic size and the valence vary within the periodic table.
Some columns are useful, I, II, VIII, VII, perhaps VI. That's at least half of the main groups. Other parts are not so clear. But it is possible to use the grid structure to make some approximations and predictions of the properties.
I wonder if it is possible to find much more general molecules that might have interesting properties through ML.
Original paper: https://journals.aps.org/prl/pdf/10.1103/PhysRevLett.117.135...