In order to elicit contributions (read Contributing), we outline below the key elements of our roadmap, which is currently under development. Some modules and packages produced within the framework of this roadmap are already present in the current GEMSEO version, as beta version capabilities.
Multi-fidelity MDO processes
They allow the use of multiple levels of precision for a given discipline, in order to accelerate the computational time. Basic refinement strategies will be distributed, together with switch criteria between the levels. Cooperation is welcome on other strategies.
Global optimization algorithms, based on deterministic as well as heuristic approaches, are being interfaced with GEMSEO. Surrogate-based optimization capabilities are also under development. A connection with the Cuter/Cutest problem test suite will be published.
GEMSEO proposes a machine learning (ML) package to easily interface any ML algorithm, for surrogate-based modeling in the presence of expensive or noisy codes. We are gradually moving towards adaptive learning methods using infill criteria, for surrogate-based optimization, MDO and uncertainty quantification.
The contributors are welcome to make use of GEMSEO concepts flexibility and develop new MDO formulations. The multi-level formulations family seems in particular very promising and the GEMSEO team would be happy to support an active collaborative community on the subject.
An extension of the GEMS process classes (MDODiscipline, MDAs, MDOChain etc), with the adjoint capabilities, to MPI communications will be released. We are open to collaborate on real applications using these features, especially high fidelity HPC ones.
Uncertainty quantification and management
GEMSEO already provides OpenTURNS-based statistical and probabilistic capabilities to describe uncertainty sources, propagate them in a multidisciplinary system and quantify the resulting output uncertainty. We are currently applying these basic functionalities to connect GEMSEO to different UQ&M fields, such as sensitivity analysis, robust optimization and robust MDO.
ODE solvers with adjoint capabilities are being interfaced with GEMSEO. The Petsc TSAdjoint library is targeted; and we are open to collaboration on that subject.
Scalable data-driven models
The contributors are welcome to evaluate and further develop this methodology. Possible extensions could be to include cross-effects in the learning stage or to make machine learning models scalable.
Benchmark MDO problems
Our roadmap also aims at enriching the list of existing MDO benchmarking problems with new ones and further assess comparative advantages of the MDO formulations. We are happy to share these elements with the MDO community and are open to collaborate on the creation of benchmark MDO problems. In addition, a general framework to compare drivers, formulations, or any algorithm is under development.
Finally we are convinced that interoperability between platforms and methodologies is key. One objective is to interoperate GEMSEO with OpenMDAO and WhatsOpt. The support of the MDO community is welcome.