Event Information:

• Thu

  05

  Nov

  2026

  SC Seminar: Bora Kalpakli

  10:15Hybrid (Room 32-349 and via Zoom)

  Dr. Bora Kalpakli, LIKUA, Heidelberg, Germany

  Title: NeuralFlow: Learning CFD Solution Operators Through Solver Backpropagation
  

  Abstract:

  Computational Fluid Dynamics (CFD) traditionally obtains flow solutions by repeatedly
  applying numerical operations that reduce the imbalance in the discretized flow field until
  a converged state is reached. Although modern CFD algorithms are accurate and robust,
  their computational cost remains a limitation for applications requiring repeated simulations,
  large parameter studies, optimization, or near-real-time prediction. NeuralFlow investigates
  a different approach: a neural operator is trained not to reproduce a database of previously
  computed CFD solutions, but to learn how the CFD solver itself drives a flow field toward
  convergence.
  NeuralFlow is based on a Graph Neural Operator (GNO) defined directly on an unstructured
  finite-volume mesh. Mesh cells are represented as graph nodes and their physical
  connections as edges. Starting from an initial flow field, the GNO predicts corrections to the
  primitive variables and is applied recurrently over multiple rollout steps. This local graph
  formulation allows the same learned operator to act on different parts of an unstructured
  mesh while progressively transforming the current state toward a converged flow solution.
  A central feature of NeuralFlow is that no reference CFD solution data are required for
  training. It is also fundamentally different from a Physics-Informed Neural Network (PINN):
  the governing differential equations are not imposed directly on the neural network as a
  collocation-based loss. Instead, the predicted flow state is evaluated by the existing CMPS
  CFD solver. CMPS performs its normal finite-volume operations, including numerical flux
  evaluation, boundary-condition treatment, residual assembly, and Jacobian evaluation. Exact
  solver sensitivities are obtained by automatic differentiation and are used to construct the
  gradient with respect to the GNO output. This solver-derived gradient is then backpropagated
  through the neural network, allowing the GNO weights to learn directly from the behaviour
  of the numerical CFD algorithm.
  The seminar will present the NeuralFlow architecture, the graph representation of finitevolume
  meshes, recurrent rollout training, and the coupling between forward-mode automatic
  differentiation inside the CFD solver and reverse-mode backpropagation inside the neural
  network. Initial results show that the learned operator can progressively reduce CFD
  residuals and reproduce laminar-flow behaviour without supervised solution data. The longterm
  objective is a generalizable learned CFD solver that retains the numerical and physical
  structure of conventional CFD while transferring part of the iterative solution process to a
  trainable graph neural operator.

  How to join online

  You can join online via Zoom, using the following link:
  https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1