The challenges in design optimisation usually consist of solving models with a very large number of decision variables and parameters that are subject to high-dimensional uncertainty that manifests itself over multiple time scales. Selected optimisation problems in aerodynamics, fluid and structural mechanics have been addressed successfully with adjoint-based optimization techniques, often leading to surprising designs with superior performance. With so-called adjoint techniques, the sensitivity of design objectives and constraints towards all design parameters are computed with a single additional simulation, providing extremely valuable information at a low cost. Thus far, these approaches have mostly been applied to comparatively simple systems. In practice, design problems often consist of complex, multi-physics simulations with intricate couplings between subsystems and a large number of engineering constraints. The design parameters can be geometric or process control parameters, which are usually required for real-time control of experiments. An example is the use of plasma edge models in the design process of critical components, the plasma facing components (PFCs) to withstand the plasma thermal loads in a nuclear fusion reactor (tokamak). Adjoint approaches can be extremely valuable in disentangling the complex, hidden dependencies, when coupled with artificial intelligence (AI) methods such as deep neural networks (DNN) that require sweeping through the parameter space for optimal plasma control parameters. In tokamak plasma edge modelling and divertor monoblock design, adjoint-based optimization has been introduced, based on simplified models. These methodologies show great promise towards improved divertor designs and model calibration, but 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 to handle the interrelated sensitivities between multi-physics modules and constraints. The development of such optimization framework and the necessary numerical tools is the central objective of this project.
Research projects are (co)financed by the Slovenian Research and Innovation Agency
- Member of the University of Ljubljana: Faculty of Mechanical Engineering
- Project code: J2-60053
- Science: Engineering sciences and technologies
- SICRIS: Adjoint and AI optimisation framework for constrained optimization with complex, multi-physics simulation models (A2FOMS)
Reporting period M1-M12 (january 2025-december 2025)
1.Objective: To develop a robust and automated network generation system and an integrated sensitivity framework. Development of automated tools for generating networks and expanding sensitivity for complex plasma edge simulations.
In the first year of the project, research activities focused on mathematical modelling and computational approaches relevant to complex simulation-based optimization problems. Special emphasis was placed on the theoretical analysis of plasma processes and mathematical descriptions of plasma dynamics, which are important for numerical modeling of fusion plasmas. These studies represent an important theoretical basis for the development of automated optimization frameworks for plasma edge simulations. Recent work has looked at analytical approaches to describe plasma behavior in kinetic systems and quasi-linear modeling techniques that contribute to a better understanding of plasma dynamics in numerical simulation environments [1].
These developments support the long-term goal of the A2FOMS project, which aims to combine advanced modelling approaches with automated optimisation workflows in plasma edge simulation tools such as SOLPS-ITER. The obtained results form the basis for improved modeling capabilities necessary for the efficient optimization of fusion reactor components.
[1] Leon Kos and Davy D. Tskhakaya. On Problems Solved in a Quasi-Linear Approximation, Mathematics. COBISS. SI-ID 272477699.
Objective 2: To develop an efficient framework for adjunctive optimization in complex multi-physical workflows. Development of an optimization framework that combines adjoint sensitivity analysis with modern computational techniques.
Research in the first year of the project focused on computational approaches to improve the efficiency of large plasma simulations, which are important for optimization workflows. In particular, machine learning techniques were studied to accelerate particle simulations in the cell in the tokamak removal layer. The developed approach uses segmented surrogate models to approximate the computationally demanding components of the simulation, which significantly reduces computational costs while maintaining the physical accuracy of the model [2]. These developments are directly relevant to the objectives of the A2FOMS project, where the efficient optimization of complex multiphysical simulation models requires fast and reliable numerical solvers. The proposed approach to machine learning represents an important methodological step towards incorporating data-driven acceleration techniques into optimization workflows used in plasma edge modeling and fusion reactor design.
[2] Nikola Vukašinović, Uroš Urbas, Leon Kos, and Ivona Vasileska. Accelerating Particle-in-Cell simulations in Tokamak Scrape-off Layer using segmented surrogate models. COBISS. SI-ID 270229507.
Objective 3: Automated treatment of complex constraints of nonlinear design and state in multi-physical design problems. Development of methods for incorporating realistic engineering constraints into optimization workflows.
In the first year of the project, research activities also addressed the mathematical and computational aspects of nonlinear dynamical processes that are important for plasma modeling. Special attention was paid to analytical studies of nonlinear wave-particle interactions and related plasma processes. These studies contribute to a deeper understanding of the physical mechanisms that influence plasma behavior and therefore play an important role in the development of reliable numerical simulation tools [3].
The obtained results provide valuable insights into nonlinear plasma dynamics and represent an important contribution to modeling approaches used in edge plasma simulations. A better understanding of these processes supports the development of optimization methods that are able to cope with complex physical constraints in realistic plasma modeling scenarios.
[3] Leon Kos, Ivona Vasileska, and Davy D. Tskhakaya. On the Theory of Nonlinear Landau Damping, Symmetry. COBISS. SI-ID 238548995.
4.Objective: To test the optimization tools in SOLPS-ITER. Testing and validation of the developed optimization tools in realistic plasma modeling environments.
In the first year of the project, methodological work focused on theoretical and computational studies relevant to plasma modeling and numerical simulation methods. These studies provide an important foundation for the future implementation and testing of optimization tools within plasma edge simulation frameworks.