A study on dynamic active learning for meta-modelling of process simulations

Accurate simulations are of fundamental importance for process engineers, but most modern simulation software is often inflexible, and cannot be easily integrated into new applications. Starting from this motivation, we propose the use of Dynamic Active Learning, the extension of active learning algorithms to dynamic processes simulations in order to generate meta-models which can be flexibly implemented into any application . The algorithm is described, implemented and tested on three process simulations (generated with Aspen Hysys), conducting sensitivity experiments on the main implementation parameters and the complexity of the simulations. The results are compared with a benchmark of (optimally chosen) uniform random inputs, showing an increase in the explored volume on the domain of the process function, which goes up to 120% for the more complex simulation, with an increase in the time demand of only 30% . The results also highlight the importance of the choice of sampling method, since in all cases the space exploration gains quickly diminish as the number of samples increase.