Prior pruning of the reaction model

The practice of SM with chemical kinetics models showed that not all active variables can be determined by optimization, as the system as a whole is under-constrained. In other words, the objective function is not sufficiently sensitive to some of the optimization variables—those that are in the tail of the screening sensitivity plot. The low sensitivity manifests itself in the appearance of the valley in the objective function, as discussed in the beginning of this chapter, and this geometric feature makes it difficult to determine numerically the global minimum, and the 

There are practical implications that follow from this observation. First, the assignment of the threshold level for separation of the model parameters into active and inactive variables becomes less critical than might appear at first. There is a definite zone of comfort in making this initially appearing as arbitrary decision if too many variables are selected to be active at the sensitivity-ranking step, the number of optimization variables can be adjusted at a later stage without any harm done. 

If we know that only high-sensitivity model parameters end up playing role in model optimization, then we may (and should) pre-screen the reaction model prior to identification of active variables and construct a smaller, more praetical model. This brings us to the discussion of construction and completeness of chemical reaction models. Fifteen years ago, the following prospective was suggested [45]  

There is a current effort to create such a data infrastructure, called Process Informatics Model (PrIMe), with the initial focus on gas-phase chemical reaction kinetics [46]. Starting with the assumption that such a data infrastructure exists, the following strategy is fostered by the framework of SM [47]. 

The initial reaction model, referred to as the first-trial model, is assembled to contain as many reactions as perceived necessary to describe the physical reality of a given response. The consideration at this stage should be given to the completeness of the reaction set, including rather than excluding possible reactions if in doubt, without concern over the reaction set size. The completeness of the first-trial reaction set should be understood as pertained to a single response, not all available or possible ones. In other words, in principle, one constructs a first-trial model for each individual response of the training set. 

---