Coupled Discrete Adjoints for Multiphysics in SU2

preCICE v2: A sustainable and user-friendly coupling library

“…Similarly to the adapter, SU2 is also used as a solver in other coupling projects, for example CUPyDO 79 , where the CUPyDO coupler calls SU2 via a Python wrapper. In addition to external coupling options, SU2 offers monolithic capabilities for multi-physics simulation such as FSI and CHT 80,81 .…”

Pitfalls of Discrete Adjoint Fixed-Points Based on Algorithmic Differentiation

“…The implication is that if iterative processes within the adjoint (that arise from differentiating M) are not converged to reasonable accuracy, even the asymptotic convergence behavior will depend on , the right-hand side of the adjoint equation (5). Furthermore, in partitioned multiphysics contexts, this right-hand side includes contributions due to the coupling between different solvers [4,9], which can be less smooth than the gradient of the objective function depending on the type of interface interpolation used. Empirically, this translates into linear system residuals of (10 −4 ) being required by this type of adjoint solver, whereas (10 −1 ) often suffices for the primal solver.…”

“…The second class of discrete adjoint methods use a more generic approach, that, in principle, does not require any information about the primal solution, and typically makes extensive use of algorithmic differentiation (AD). The generality of AD is especially appealing in multiphysics contexts to allow the development of easily extendable design and simulation software [4].…”

Aerodynamic Shape Optimization Based on Discrete Adjoint and Rbf

“…The Jacobians involved in the formula (15) are obtained automatically differentiating different the SU2 code using the library CoDiPack. A detail description can be found in [10]. In practice, to obtain the evaluation chain, all the variables are substitute by a new datatype that stores at the same time the value and the elementary operations involved.…”