Table 2. Overview of Learning Methods That We Discussed in Section 5 and Examples of Their Use in the Field of Porous Materials^a.

method	section	application to porous materials
representation learning
HDNNP	5.1.1.1	trained on fragments for MOF-5 by Behler and co-workers286
message-passing NN	5.1.1.2	not used for porous materials so far
convolutional or recurrent NN	5.1.1.3	Wang et al. used CNN to classify MOFs based on their XRPD pattern135
crystal-graph based models	5.1.1.6	Korolev et al. use them to predict bulk and shear moduli of pure silica zeolites and Xe/Kr selectivity of MOFs433
generative models	2.1.2.2.2	ZeoGAN by Kim and co-workers434 (cf. section 9.7)
classical statistical learning
linear models	5.2.1	predicting gas uptakes based on tabular data of simple geometric descriptors246
kernel methods	5.2.2	predicting gas uptakes based on graphs and geometric properties,435 might be also interesting in the SOAP-GAP framework, as work by Ceriotti and co-workers as well as Chehaibou et al. showed436,437
ensemble models	5.2.5	often used in form of RF or GBDT to predict gas uptakes based on tabular data of simple geometric descriptors, ensemble used to estimate uncertainty when predicting oxidation states438
Bayesian methods	5.2.3	have been used, e.g., in the form of GPR435 or Bayesian NN439,440 but not all features, like the uncertainty measure, have been fully exploited so far. This might be useful for active learning, e.g. for MD simulations in the Bayesian formulation of the SOAP-GAP framework
TDA	4.2.2.4.1	Moosavi, Xu et al. built KRR models for gas uptake in porous organic cages,185 or Zhang et al. for gas uptake in MOF,233 Lee et al. for similarity analysis237
other ML techniques
automated machine learning	10.1	Tsamardinos et al.441 use the Just Add Data tool to predict the CH4 and CO2 capacity of MOFs, Borboudakis et al. use the same tool to predict CO2 and H2 uptakes172
data augmentation	3	Wang et al. used it for the detection of MOFs based on their diffraction patterns135
transfer learning	10.3	He et al. used it for the prediction of band gaps245
active learning	3.3	could be used for MD simulations using ML force fields,312 or to guide the selection of next experiments or computations
capturing the provenance of ML experiments	10.2	Jablonka et al. used comet.ml to track the experiments they ran for building models that can predict the oxidation state of metal centers in MOFs438
Δ-ML	10.3	Chehaibou et al. used a Δ-ML approach to predict random phase approximation (RPA) adsorption energies in zeolites437

^aFor some methods there has been no application reported in the field of porous materials, and we instead provide ideas of possible applications.