The objective of Deliverable 5.2 (D5.2) consists on the development of Reduced Order Models (RoMs) derived from the comprehensive full-order models presented in the previous deliverable, D5.1. The approach presented in this deliverable involved the construction of a robust dataset through a systematic design of experiments (DoE), by means of carrying out simulations under different combinations of four distinct process variables. A total of 300 trials were carried out under varying conditions, ensuring a substantial foundation, and providing plentiful enough data for constructing the RoMs.
Strategically enhancing computational efficiency, a parallelization strategy was devised and integrated into the trial execution process. This approach yielded significant reductions in computational time, underscoring the advantage of leveraging the open-source FVM library fipy. This open-source framework provided an elevated degree of flexibility compared to conventional commercial software, thereby allowing comprehensive engagement with the full-order models.
The subsequent dataset, assembled using the aforementioned approach, served as the substrate for training an array of machine learning (ML) algorithms of diverse natures. These algorithms were trained to fit the four distinct outputs under consideration. To address scenarios where a single-output model fell short, different workarounds were conceived and applied to enable the production of multiple outputs using them.
Training of the models was undertaken, accompanied by the formulation of specific optimization functions for each one. This process involved the exploration of tailored hyperparameter spaces to derive the near-optimal model configurations. Various model associations, including a chain of regressors, were also evaluated in addition to the use of those consisting in single layer regressors within the model pipelines.
Upon exhaustive analysis, a synergistic ensemble of Support Vector Regressors (SVRs) demonstrated superior performance in contrast to alternative model configurations. Consequently, this amalgam of SVRs was identified as the most suitable RoM for integration within the CORALIS platform.
By harmonizing design of experiments, computing parallelization, advanced machine learning techniques, and methodical model optimization, deliverable 5.2 culminated in the selection of an optimal RoM solution, poised to augment the capabilities of the CORALIS platform and contribute to the overarching project objectives.
Furthermore, it’s worth highlighting that we are currently in the process of developing a Computational Fluid Dynamics (CFD) model for the denitrification process subsequent to this procedure. Anticipated future enhancements to this model are expected to be informed and improved through the incorporation of the findings from this work.