SoftCombi DoE Tool

For Catalyst Development


Click to enlarge

The proposed soft computing framework combines a genetic algorithm (GA) and an artificial neural network (ANN). Thus, ANN is employed as an approximated model for fitness evaluation, whereas the developed GA tries to find the optimal solution by investigating several catalysts simultaneously. Specifically, this soft computing technique (see Figure) consists of the following steps: (i) setting up; (ii) ANN re-training; (iii) GA operators; (iv) pre-screening; and (v) experimental testing. The steps ii to iv are repeated till the convergence criteria is satisfied. However, the final actions performed in each step depend on the configuration of the tool implemented following the suggested framework. For example, it is possible to get different generations (employing fitness approximation) before carrying out the pre-screening step.

In the setting up process, the problem under study must be codified properly; next, the soft computing approach parameters are set; following, the starting generation of the optimization process is calculated; and finally a suitable ANN model is obtained for
predicting the catalytic performance.

Regarding the ANN re-training step, new experimental data derived from the testing of each succeeding generation is divided into training and testing data. The training set is used to retrain the stored ANN, whereas the testing set is employed to compare the stored ANN and the newly-retrained ANN. So that one with the best predicting performance is selected and stored.

The candidates for the new generation are designed by the GA operators, taking into account the previous experimental results. Specifically, the GA employs mutation and crossover operators, requiring the last one the assistance of the ANN.

The pre-screening process is done in two phases. Firstly an approximation fitness value for each sample is calculated by means of the ANN predictions. Secondly, a controlled generation is obtained from the virtual population, reducing the number of samples according to the value of the reduction ratio parameter.

Finally, the pre-screened generation is experimentally tested, getting the controlled fitness evaluations of the last generation proposed by the GA. This controlled values are updated in the system and later, they are employed in the next optimization loop.