Learning differentiable simulation models of control systems
Time: Fri 2026-06-05 09.00
Location: F3 (Flodis), Lindstedtsvägen 26 & 28, Stockholm
Video link: https://kth-se.zoom.us/j/65571659755
Language: English
Subject area: Electrical Engineering
Doctoral student: Miguel Aguiar , Reglerteknik
Opponent: Giancarlo Ferrari-Trecate, École polytechnique fédérale de Lausanne, Lausanne, Switzerland
Supervisor: Professor Karl H. Johansson, Reglerteknik; João Borges de Sousa, University of Porto, Porto, Portugal; Amritam Das, TU Eindhoven, The Netherlands
QC 20260508
Abstract
The use of conventional model-based frameworks for design of estimation and control algorithms is precluded by the increasing complexity of engineering systems. To address this challenge, in this thesis we study new paradigms for data-driven modelling of dynamical systems based on deep learning methods. Three contributions are developed on fundamental algorithms and on applications of the techniques.
Our first and main contribution is a method for learning continuous-time simulation models of control systems. The method uses a novel neural network architecture to approximate the solution operator of such systems. The resulting model enables efficient simulation of trajectories, and is differentiable, meaning one can easily compute gradients of trajectories with respect to initial conditions, system parameters, and external inputs. These properties are desirable for model-based optimisation tasks such as parameter estimation, and we show that our method is highly effective in such contexts. We consider surrogate modelling of excitable systems, provide an extension to spatiotemporal control systems, and prove that the architecture is a universal approximator of flows of a general class of control systems.
In the second contribution we develop physics-informed learning methods for partial differential equations. We propose a new method that addresses the spectral bias problem, with performance superior to the state of the art on several baselines. In an application to traffic systems, we provide a method to learn solutions and equation parameters when measurements are sparse.
The third contribution is on the use of learning-based simulation models in unmanned vehicle operations for ocean observation, specifically for mission planning, adaptive sampling, and trajectory planning, in which simulation of dynamical systems plays a key role. We develop algorithms for sampling ocean fronts using a single vehicle, showing in numerical and software-in-the-loop simulations that we are able to track fronts using prior data. We propose mission and trajectory optimisation methods for multi-stage missions with multiple vehicles, developing an efficient dynamic programming algorithm with an associated software architecture.