Troubleshooting

Experience has shown that the first few attempts at running PROCESS with a new file tends to produce infeasible results; that is, the code will not find a consistent set of machine parameters. The highly non-linear nature of the numerics of PROCESS is the reason for this difficulty, and it often requires a great deal of painstaking adjustment of the input file to overcome.

Error handling

In general, errors detected during a run are handled in a consistent manner, with the code producing useful diagnostic messages to help the user understand what has happened.

There are three levels of errors and warnings that may occur:

• Level 1 -- An informational message is produced under certain conditions, for example if the code modified the user's input choice for some reason.
• Level 2 -- A warning message is produced if a non-fatal situation has occurred that may result in an output case that is inaccurate or unreliable in some way.
• Level 3 -- An error message will occur is a severe of fatal error has occurred and the program cannot continue.

These messages are printed on the screen during the course of a run, and those still active at the final (feasible or unfeasible) solution point are also written to the end of the output file (messages encountered during the iteration process are not copied to the output file, as the convergence to a valid solution might resolve some of the warnings produced earlier in the solution process).

The error_status variable returns the highest security level that has been encountered (or zero if no abnormal conditions have been found); of a severe error (level 3) is flagged at any point the program is terminated immediately. The final message number encountered during a run is returned via output variable error_id . In addition, with certain messages, a number of diagnostic values may also be given; these can be used to provide extra diagnostic information if the source code is available

General problems

A code of the size and complexity of PROCESS contains myriads of equations and variables. Virtually everything depends indirectly on everything else because of the nature of the code structure, so perhaps it is not surprising that it is often difficult to achieve a successful outcome.

Naturally, problems will occur if some of the parameters become un-physical. For example, if the aspect ratio becomes less than or equal to one, then we must expect problems to appear. For this reason, the bounds on the iteration variables should be selected with care.

Occasionally arithmetic ("NaN") errors are reported. They usually occur when the code is exploring un-physical values of the parameters, and often suggest that no feasible solution exists for the input file used.

The error messages produced by the code attempt to provide diagnostic information, telling the user where the problems occurs, and also suggest a possible solution. These messages are out of necessity brief, and so cannot promise to lead to a more successful outcome.

The is the option to turn on extra debugging output; to do this, set verbose = 1 in the input file.

Optimisation problems

On reflection it is perhaps surprising that PROCESS ever does manage to find the global minimum figure of merit value, if there are nvar iteration variables active the search is over nvar-dimensional parameter space, in which there may be many shallow minima of approximately equal depth. Remember that nvar is usually of the order of twenty.

The machine found by PROCESS may not, therefore, be the absolute optimal device. It is quite easy to have two or more solutions, with results only a few percent different, but a long way apart in parameter space. The technique of "stationary" scans is sometimes used in this situation: a scan is requested, but the same value of the scan variable is listed repeatedly.

Scans should be started in the middle of a range of values, to try to keep the scan within the same family of machines. The optimum machine found may otherwise suddenly jump to a new region of parameter space, causing the output variables to seem to vary unpredictably with the scanning variable.

It should be noted that in general the machine produced by PROCESS will always sit against one or more operation limits. If, during a scan, the limit being leant upon changes (i.e. if the machine jumps from leaning on the beta limit to leaning on the density limit) the output parameters may well become discontinuous in gradient, and trends may suddenly change direction.

Unfeasible results

In the numerics section of the output file, the code indicates whether the run produced a feasible or unfeasible result.

The former implies a successful outcome, although it is always worth checking that the sum of the squares of the constraint residuals (sqsumsq) is small ($~10^{-3}$ or less); the code will issue a warning if the solver reports convergence but the value of sqsumsq exceeds $10^{-2}$. If this occurs, reducing the value of the HYBRD tolerance ftol or VMCON tolerance epsvmc as appropriate should indicate whether the result is valid ot not; the output can usually be trusted of (1) the constraint residuals¹ fall as the tolerance is reduced to about $10^{-8}$, and (2) the code indicates that a feasible solution is still found.

An unfeasible result occurs if PROCESS cannot find a set of values for the iteration variables which satisfies all the given constraints. In this case, the values of the constraint residues shown in the output give some indication of which constraint equations are not being satisfied - those with the highest residues should be examined further. In optimisation mode, the code also indicates which iteration variables lie at the edge of their allowed range.

Unfeasible runs can be caused by specifying physical incompatible input parameters, using insufficient iteration variables, or by starting the problem with unsuitable values of the iteration variables.

The utility run_process carries out many runs, changing the starting values of the iteration variables randomly. It stops once a feasible solution is found.

It is important to choose the right number of useful iteration variables for the problem to be solved - it is possible to activate too many iteration variables as well as too few, some of which may be redundant.

Both optimisation and non-optimisation runs can fail with an error message suggesting that the iteration process is not making good progress. This is likely to be due to the code itself unable to escape a region of the parameters space where the minimum in the residuals is significantly above zero. In this situation, there is either no solution possible (the residuals can therefore never approach zero), or the topology of the local minimum makes it difficult for the code to escape to the global minimum. Again, a helpful technique os to wither change the list of iteration variables in use, or to simply modify their initial values to try to help the code avoid such regions.

A technique that occasionally removes problems due to unfeasible results, particularly if an error code ifail = 3 is encountered during an optimisation run, is to adjust slightly one of the limits imposed on the iteration variables, even if the limit in question has not been reached. This subtly alters the gradients computed by the code during the iteration process and may tip the balance so that the code decides that the device produced is feasible after all. For instance, a certain component's temperature might be 400 K, and its maximum allowable temperature is 1000 K. Adjusting this limit to 900 K (which will make no different to the actual temperature) may be enough to persuade the code that it has found a feasible solution.

Similarly, the order in which the constraint equations and iteration variables are stored in the icc and ixc arrays can make the difference between a feasible and unfeasible result. This seemingly illogical behaviour is typical of the way in which the code works.

Another technique in such situations may be to change the finite difference step length epsfcn, as this might subtly change the path taken in the approach towards a solution.

It may be the case that the act of satisfying all the required constraints is impossible. No machine can exist if the allowed operating regime is too restrictive, or if two constraints require conflicting (non-overlapping) parameters spaces. In this case some relaxation of the requirements is needed for the code to produce a successful machine design.

Hints

The above sections should indicate that it is the complex inter-play between the constraint equations and the iteration variables that determines whether the code will eb successful at producing a useful result. It can be somewhat laborious process to arrive at a working vase, and experience is often of great value in this situation.

It should be remembered that sufficient iteration variables should be used to solve each constraint equation. For instance, a specific limit equation may be $A \leq B$, i.e. $A = fB$, where the f-value $f$ must lie between zero and one for the relation to be satisfied. However, if none of the iteration variables have any effect on the values of $A$ and $B$, and $A$ happens to be greater than $B$, the PROCESS will clearly not be able to solve the constraint.

The lower and upper bounds of the iteration variables are all available to be changed in the input file. Constraints can be relaxed in a controlled manner by moving these bounds, although in some cases care should be taken to ensure that un-physical values cannot occur. The code indicates which iteration variables lie at the edge of their range.

It is suggested that constraint equations should be added one at a time, with required new iteration variables activated at each step. If the situation becomes unfeasible it can be helpful to reset the initial iteration variable values to those shown in the output from a previous feasible case and rerun the code.

---

1. The constraint residuals are the final values of $c_i$ in the constraint equations. The value sqsumsq is the square root of the sum of the squares of these residuals. ↩