Parametric methods Parametric methods Parametric methods Parametric methods Semiparametric Nonparametric Nonparametric
Lumping Conservation Analysis Time scale separation Sensitivity analysis Balanced Truncation Artificial neural network Empirical approximation
System agnosticism Utility depends on the nature of the full-order model to be reduced. Not all models are amenable to these methods Utility depends on the nature of the full-order model to be reduced. Not all models are amenable to these methods Utility depends on the nature of the full-order model to be reduced. Not all models are amenable to these methods Utility depends on the nature of the full-order model to be reduced. Not all models are amenable to these methods Have no requirement about the nature of the full-order model Have no requirement about the nature of the full-order model Have no requirement about the nature of the full-order model
Automatability Yes Yes Some variants are automatable Yes Yes Yes No
Mechanistic relevance Yes Yes Yes Yes No No No
Accuracy Determined by the user as a trade-off of complexity vs accuracy No loss of accuracy incurred- but only minimal model-order reduction is possible with this technique Variable, depending on the system Variable, depending on the system Determined by the user as a trade-off of complexity vs accuracy Accuracy can be controlled by the size of the network Can be any level of accuracy required but only for 1 or 3 response variables
Computational ease Computationally demanding for large nonlinear models Efficient Variable Computationally demanding for large nonlinear models Efficient, although upfront cost of simulating pseudo-data may be high in some situations Very efficient, limited data needed
Predictive ability of the reduced model (interpolation and/or extrapolation) Both Retains all properties of the full-order model including extrapolation ability Both Both Possibly, Can be used to describe data based on the input and output functions used to generate the reduced model Can only predict within the range of pseudo-data (interpolation). Not useful for extrapolation Can only predict within the range of pseudo-data (interpolation). Not useful for extrapolation
Requirement of experimental data for reduced model parameter estimation No No No No No Yes – needs data or pseudo-data to create model structure Yes – needs data or pseudo-data to create model structure
Suitable for parameter estimation Yes Yes Yes Yes No No Yes