Simanopt is a process simulation and optimization system built around a set of very fast simulation models. The models employ simplified physical properties that enable analytical or semi-analytical solutions. The Chemcept-data physical properties system is employed to provide the required physical properties. The principles of the fast reliable simulation are described in the pages on physical properties . Essentially, we are aiming for simulations ten thousand times faster than conventional process simulations. At the same time, we are aiming at extreme reliability. Work published by Johns & Vadhwana has shown that these goals are achievable. There is no benefit for a one-off simulation. For such simulations, the extra precision available from a conventional simulator outweighs any speed disadvantage. Furthermore, for one-off simulations, the user usually has a good estimate of the solution. This estimate permits good starting estimates for iteration so that extreme model reliability is not required.

For most applications, it is not worth investing time to make much faster computer run times. Certainly, speed increases of a factor of two to five are quickly overtaken by advances in computer hardware. A factor of two to five is the most that can be achieved by conventional code optimization. A speed increase of 10000 requires a radically different approach. It provides a speed advantage that will require 25 to 30 years of hardware advances to match. Furthermore, when coupled with improved reliability, it permits a range of applications that are otherwise scarcely practicable. Simanopt has the following application areas:

• 1) Very fast, reliable simulation. This feature is, in itself, of marginal benefit. However, it enables results to be obtained for difficult problems that may crash a conventional simulator.

• 2) Optimization. Multi-variable optimization may take thousands of iterations to converge. Acceptable conventional run times of a fraction of a minute become unacceptable run times of a fraction of a day. Furthermore, during the iteration, the optimizer may throw completely unrealistic values at the simulator (columns with millions of stages, pressures of millions of bars etc). These unrealistic values will be eliminated as the optimization progresses. However, they may crash a conventional simulator and prevent the optimal solution ever being reached. Often a conventional simulator will be used to fine-tune the optimum. However, even in these cases, the excellent initial estimate provided by Simanopt converts an impossible optimization to a feasible optimization.

• 3) Simulation under uncertainty. Most engineering designs are undertaken in the face of uncertainty. Statistical information may be available on the data uncertainties such as physical properties, model performance. However, thousands of simulations may be required to estimate performance statistics from data statistics. Again, a very fast simulator is required. Any additional uncertainty introduced by the simplified models can be scoped. Thus, the simulation itself can estimate the potential impact of the simplified model uncertainties compared to the inevitable data uncertainties. The results provide guidance on whether further simulation with a conventional high-quality simulator is needed.

• 4) Optimization under uncertainty. These optimizations are challenging and may require tens of thousands of model simulations. They are particularly challenging when the optimization includes conditional decisions. For example, we may create a "nest" such as:

  Optimize (generate potentially optimal values of adjustable design variables)
       Evaluate Uncertainty (generate possible outcomes for each uncertain variable)
            Re-optimize (generate potentially optimal values of operating variables)
  

  This type of optimization strategy is realistic. Thus, in the outermost loop, we generate possible reactor sizes, separator sizes etc. In the next loop, we generate possible outcomes for the physical design generated in the outer loop. For example, we may be uncertain of a reaction rate. Possible outcomes would be that the reaction rate is high, or the reaction rate is low. If the rate is high, there is an operation optimization in the inner loop. For example, we could increase production rate, decrease recycle of unconverted raw materials, or increase product purity. If the rate is low, we may reduce production rate, increase recycle of unconverted raw material, or pay the cost penalty of making lower specification product. Consider the situation in which, for every set of uncertain outcomes, it takes 100 iterations to optimize the operating variables. Consider also that for every set of design variables, we need to generate 100 sets of possible outcomes adequately to cover the range of uncertainties. Finally, consider that it takes 100 iterations to optimize the design variables. Under these conditions, it takes 100×100×100 process simulations to reach an optimal design. Very fast, reliable simulation is required to get results in any reasonable computational time.

• 5) Process Synthesis. In Process Synthesis, we augment conventional optimization to include integer variables. Thus, in a conventional optimization, we optimize lengths, diameters and flow rates, which are all real numbers. Conventional non-linear optimization handles such problems. In Process Synthesis, we also make optimal decisions about technology and connectivity. Technology decisions may include "do we use distillation or liquid/liquid extraction?". Connectivity decisions may include "do we recycle unconverted raw material, if so, to where do we recycle it?". These decisions are included as additional variables. They are normally binary variables, such as: 0 = distillation, 1 = liquid/liquid extraction. Adding these binary variables increases computer run times exponentially. Thus, the simplest way to handle a problem with 10 binary variables would be to consider every possible combination of the variables, giving 1024 combinations. We then have 1024 conventional optimizations. Efficient integer optimization techniques are available that require only a fraction of these combinations to be evaluated. Furthermore, they do not require complete optimization of the real-value subproblems in order to identify the optimum. Nevertheless, it can be proven that computer run times still increase exponentially with number of binary variables. At the innermost loop, the process may be simulated millions of times. Very fast simulation is required to handle such problems. Furthermore, very robust simulations are required. As structures are changed automatically, completely unrealistic values may be thrown at the simulator for which realistic results are required.

Note that the Simanopt approach becomes quite impractical for large numbers of integer variables. It also requires a prior choice as to what integer options might be required. For problems with very large numbers of integer variables, and in which the possible process structures cannot be defined in advance, our other complementary tool, Chemcept-Design is the tool of choice.