Adjoint and AI optimisation framework for constrained optimization with complex, multi-physics simulation models

Financiers

Challenges in structural optimization usually involve solving models with a very large number of decision variables and parameters, which are subject to high-dimensional uncertainty manifesting across multiple time periods. Selected optimization problems in aerodynamics, fluid mechanics, and structural mechanics have been successfully solved using adjoint optimization techniques, which often lead to surprising models with outstanding performance. With these so-called adjoint methods, the sensitivity of design objectives and constraints to all design parameters can be computed with just a single additional simulation, providing extremely valuable information at low cost. Until now, these approaches have mostly been applied to relatively simple systems. In practice, however, design problems often consist of complex, multiphysics simulations with intricate couplings between subsystems and a large number of engineering constraints.

Design parameters may include geometric features or process control parameters, which are typically required for real-time control in experiments. An example is the use of plasma edge models in the design process of critical plasma-facing components (PFCs) that must withstand plasma heat loads in a nuclear fusion reactor (tokamak). Adjoint approaches can be extremely valuable in disentangling complex hidden dependencies when combined with artificial intelligence (AI) methods such as deep neural networks (DNNs), which require sweeping through the parameter space to identify optimal plasma control parameters. In the modeling of tokamak plasma edges and in the design of divertor monoblocks, optimization based on aggregate effects has been introduced using simplified models. These methodologies show great promise for improved divertor design and model calibration, yet their application to realistic, constrained design problems remains a challenge. Enabling efficient adjoint-DNN optimization with realistic plasma edge models and realistic design constraints requires fundamental progress in handling interdependent sensitivities between multiphysics modules and constraints. The central objective of this proposal is to develop such an optimization framework and the necessary numerical tools.

Project Partners

ITER Organisation, Saint-Paul-lez-Durance, France

KU Leuven, Leuven, Belgium

Institute for Plasma Physics, Czech Academy of Sciences, Czechia

DTT - Divertor Tokamak Test facility, Frascati, Italy