Main Subcommittee Category: Simulation

How would you label your submission?: Industry

Abstract:

Machine Learning (ML) models have been adopted for simulation due to increased efficiency, development savings, and reduced computational requirements in production. However, there is a resistance to adopting ML due to a lack of trust in non-deterministic results that some of these algorithms produce. Crafting Digital Twins (DTs) for Defence applications is no exception, where computationally expensive Maths and Physics (MaP) based models are used to control virtual versions of real assets. Therefore, a careful blend of the nuanced understanding provided by Subject Matter Experts (SMEs) with the adaptive learning capabilities of data-driven ML solutions would be ideal to create DTs.

Physics-Informed Neural Networks (PINNs) are capable of producing dynamic models which can surpass MaP simulation performance. Similar to ML, PINNs learn from historical data to capture unique behaviours of assets as opposed to the ideal conditions MaP simulators assume. This is tempered by the influence of SMEs through physical equations which PINNs utilise as a guide in areas of the domain space where there is no training data, as well as a benchmark for data trustworthiness.

These dynamics models can be utilised by human operators or Reinforcement Learning (RL) algorithms in autonomous control systems. This approach to creating Digital Twins presents a transformative solution to rapid and cost-effective development of next-generation integrated training and readiness applications, tailored to contemporary threat environments, assuring deterrence.

This paper explores what PINNs are and their utility in creating Digital Twins. We demonstrate a practical example by training a PINN on telemetry logs of an in-service Unmanned Aerial System (UAS) and SME knowledge to create a flight dynamics model. The resulting model is then enhanced using an RL algorithm to drive the DT in a synthetic environment, concluding with an analysis of the results, discusses future certification, and proposed future work.

Keywords: AGENT-BASED SIMULATION;AI;BEHAVIOR MODELING;DEEP LEARNING;DIGITAL-GAME-BASED-LEARNING;FLIGHT SIMULATION;M&S;MACHINE LEARNING;NETWORKS;PHYSICAL MODELLING SYNTHESIS;RAPID MODELING;SYNTHETIC ENVIRONMENT;UAV

Additional Keywords not in the list above: DIGITAL-TWINS, PHYSICS-INFORMED-NEURAL-NETWORKS, REINFORCEMENT-LEARNING

Menu

Mr. Thomas McRobie