Parameterization missing parameters

Open the menu item Edit/View, there Force Field and there Bond Stretch Parameters. Move the cursor to the row C03-NT, click on this line, click on the button Delete, confirm and leave the Force field editor with OK. Now, your MOMEC force field does not have any parameters for Com-amine bonds you are ready to develop your own parameterization Actually, see what happens if you try to refine a molecule with a missing parameter Open the [Co(NH3)6]3+ - hin file from Section 17.1 (Setup/Files) or from the CSD and refine it (set the maximum number of cycles (Setup/Optimization Controls) to 2 Execute/Geometry Optimization). At the bottom of the Summary window the Force Field Messages will tell you that there is no bond stretch function for atom types C03 NT. 

There is one additional point to be mentioned in the parameterization process that is also important for understanding why the addition of missing parameters by comparison with existing data or from external sources is somewhat problematic. This is the question of redundant variables, as can be exemplified by considering acetaldehyde. 

This listing is by no means complete. Intentionally omitted from this compilation are the many papers from groups who presented quantum mechanically derived atomic charges for a molecule under study and called that a force field parameterization in the experimental section of their paper. Missing too are many publications in the literature from journals we do not read. We apologize for such omissions but confess that we cannot keep track of everything published. Nonetheless, this list will be of value to many readers of this book series. If you are aware of published parameters that are not included in the table, please forward them to the senior author. Likewise, if you find an error or an inappropriate entry, forward that information too, to ensure that the error is rectified. 

To see why this should be so, consider the effect of using a training set in which one type of reference data, geometries for example, were ab.sent. After optimization of the parameters, the minimum in the parameter hypersurface is obtained. If that minimum is well defined, then the location of the minimum must be in the same region as that for the training set in which the geometries were included. In other words, in any computational model that accurately models physical phenomena, there is only one set of values, and that set is well defined. When such a model is parameterized with some data systematically missing, and when an attempt is made to use the model to predict such data, the model is likely to work well. 

When the number of available data is small, a case can be made for using chemically meaningful constraints on the values of parameters. Under such circumstances the artificially imposed constraints or rules provide valid, useful information that would otherwise be missing. This situation occurs when elements for which there is a paucity of data (e.g., scandium) are parameterized. Numerical constraints are then likely to be necessary. 

---