Table 2 Description of revised autocorrelation (RAC) features with start/scope, operation performed, count of features removed, and total feature count.

From: MOFSimplify, machine learning models with extracted stability data of three thousand metal–organic frameworks

start

scope

operation

features removed

feature count

mc

all

product

1 (mc-I-0-all)

19

mc

all

difference

8 (Dmc-I-0-all, Dmc-I-1-all, Dmc-I-2-all, Dmc-I-3-all, Dmc-S-0-all, Dmc-T-0-all, Dmc-Z-0-all, Dmc-χ-0-all)

12

lc

linker

product

1 (lc-I-0-linker)

19

lc

linker

difference

8 (Dlc-I-0-linker, Dlc-I-1-linker, Dlc-I-2-linker, Dlc-I-3-linker, Dlc-S-0-linker, Dlc-T-0-linker, Dlc-Z-0-linker, Dlc-χ-0-linker)

12

func

linker

product

0

20

func

linker

difference

8 (Dfunc-I-0-linker, Dfunc-I-1-linker, Dfunc-I-2-linker, Dfunc-I-3-linker, Dfunc-S-0-linker, Dfunc-T-0-linker, Dfunc-Z-0-linker, Dfunc-χ-0-linker)

12

full

all

product

0

20

full

linker

product

0

20

   

26

134

  1. Five heuristic atom-wise quantities are used to perform all product and difference operations: nuclear charge (Z), electronegativity (χ), topology (T), identity (I), and covalent radius (S). MOF RACs contain four possible starts and two possible scopes: metal-centered (mc) start, linker coordinating atom centered (lc) start, functional group centered (func) start, every atom (full) start, all atom in primitive cell (all) scope, or all atom in linker (linker) scope. All starts, scopes and operations use bond depths of 0, 1, 2, and 3 to generate autocorrelations (for a total of 20 possible features for each scope). Cases that are invariant across all MOFs are listed in the “features removed” column. RAC features are given using the notation: <operation/start>-<atomic property>-<depth>-<scope>. RAC features using products are indicated by their start (e.g. mc), and those using differences contain a “D” prefix with a subscripted start (e.g. Dmc).