Scalable Framework Integration of CODA for a Multidisciplinary Preconditioned Matrix-Free Newton-Krylov Method

Solving multidisciplinary problems by coupled black-box solvers using splitting methods, such as the nonlinear block-Gauß-Seidel method, has severe limitations on convergence. We implemented a MPI-parallel full Newton-Krylov method involving the next-generation CFD solver CODA, combining the HPC platform FlowSimulator with the open-source multidisciplinary analysis and optimization Python framework OpenMDAO. The framework implementation uses the Python entry points from CODA for its residual vector, its Jacobian and its linear solution routine as matrix-free operators. For the ONERA M6 wing ( \(1.3{\cdot }10^{6}\) DoF) with torsional spring attachment, we recovered the quadratic convergence of Newton’s method and showed the reduction in computational work compared to the baseline method. The implementation is capable of handling the computationally large NASA DPW5-CRM half-plane ( \(144{\cdot }10^{6}\) DoF) distributed across a HPC cluster.