Physics-informed Artificial Intelligence for Cognitive Twins of Complex Systems
Abstract
To date, digital twins have been equipped with some intelligent features, such as data analytics, simulation, ... thus making the digital twin going beyond the pure digital replica. In this way, the asset will no longer have a digital mirror, but instead the physical and the digital versions form now a single entity with enhanced properties. To comply with the frequent and very stringent real-time requirements of digital twins, models for simulation must be of reduced order. The concept of cognitive twin has been introduced in the literature very recently. While digital twins, as digital replicas of physical assets, are nowadays common, cognitive twins represent today an evolution of this concept. We understand cognition as the ability to understand context, reason on top of existing information, predict and optimize behavior. For this to be possible, we must add artificial intelligence (AI) capabilities to the twin. Closely related, perception, the quality of being aware of things through the physical senses, especially sightÑas defined by the Cambridge DictionaryÑis another component of cognitive twins. To obtain reduced-order models of complex physical systems, we hypothesize that it is necessary to equip AI with inductive biases that should be precisely those based upon the fulfillment of known physical principles. Incorporating existing physical knowledge in the form of inductive biases seems to be the best way to (i) avoid unexpected results in previously unseen situationsÑi.e., extrapolationÑ, (ii) maximize accuracy in the predictions and (iii) minimize the amount of data necessary at the training stage. This project is aimed at giving cognitive capabilities to digital twins by using simulation as an engine of physical scene understanding. For this to be possible, we aim at incorporating physics-based artificial intelligence to their simulation modules. The following research objectives are considered: 1. The development of cognitive digital twins able to adapt themselves to changing circumstances. 2. Development of cognitive twins able to learn from ÓobservationÓ. 3. To develop twins able to work in the low data and partial data regimes. 4. Development of coupled ROMs and component-wise ROMs. This research project will lead to the development of a novel type of artificial intelligence that surpasses the state fo the art in the following aspects: - Once finished, the resulting cognitive twins will constitute a methodology able to perceive the physics of the surrounding scene. This means that the resulting physics-informed AI will be able to identify known phenomena and predict future events. - If faced to non-previously seen phenomena, this AI will be able to learn from observation, very much like toddlers do. - Information will be transmitted to humans in the loop by means of augmented reality techniques, thus easing the decision-making process. - All these predictions will comply by construction with the laws of thermodynamics, thus guaranteeing their physical significance. At the same time, such a cognitive twin will maximize the accuracy and minimize the amount of data necessary for the learning procedure. Meaningless predictions typical of black-box AI strategies are also prevented.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 20, 2022
- Source ID
- W911NF2210271
Entities
People
- Elías Cueto
Organizations
- Army Contracting Command
- United States Army
- University of Zaragoza