Back to search

NAERINGSPH-Nærings-phd

Uncertainty quantification and phenomenological knowledge: Predictive power of emulator models for tail behaviour of safety critical systems

Alternative title: Usikkerhetskvantifisering og fenomenologisk kunnskap: Presisjon av emulatormodeller for haleoppførsel av sikkerhetskritiske systemer

Awarded: NOK 1.7 mill.

Project Number:

276282

Project Period:

2017 - 2021

Funding received from:

Organisation:

Location:

Artificial Intelligence (AI) and data-driven decisions based on Machine Learning (ML) are making an impact on an increasing number of industries. As these autonomous and self-learning systems become more and more responsible for decisions that may ultimately affect the safety of people, assets, or the environment, ensuring the safe use of AI will be crucial. This project aims to provide some of the tools needed to make data-driven modeling suitable for use in safety-critical systems, like a ship, offshore structure, or a spacecraft. This is challenging when we are faced with complex physical phenomena, in environments with a high degree of uncertainty, and where the consequence of an erroneous decision can be catastrophic. To succeed, the knowledge we possess about these phenomena must be exploited optimally. We consider various ways in which knowledge about the underlying physical system can be incorporated into probabilistic models. This includes how to make use of expensive computer simulations most efficiently, and how physics-based knowledge can be used as constraints to obtain "physically obedient machine learning models". With this approach, we develop algorithms that can be used to search for optimal decisions in uncertain and safety-critical environments. The project has developed model-based approaches that exploit phenomenological knowledge to improve machine learning methods where data is scarce with respect to the complexity of the problem that is addressed. A method for making use of phenomenological knowledge as constraints in probabilistic machine learning (Gaussian Processes) has been developed. We have studied a new method for use in reliability analysis where the failure mode is unknown (Environmental Contours), resulting in new theoretical insight and an efficient algorithm for practical application. Throughout the project, we have focused on developing methods and algorithms aimed at problems related to optimal decision-making under uncertainty. We have developed an algorithm for optimal design of experiments, i.e. how to optimally collect new information, and extended this framework to a more general decision-making context using reinforcement learning. The application of such methods is especially relevant for probabilistic analysis of rare events, such as structural failure in safety-critical systems, where not enough data exist to rely on observations alone.

Vi har opparbeidet ny kunnskap som er viktig for fremtidig bruk av maskinlæring og datadreven modellering i sikkerhetskritiske systemer. Metodene som har blitt utviklet kan i dag benyttes i kommersielle prosjekter og forskningsarbeid hos DNV.

The project will seek to establish the necessary mathematical framework for introducing phenomenological knowledge in response surface models. These are fast running approximations of complex time consuming computer simulations or expensive physical experiments, based on limited realizations of the simulation/experiment and statistical machine learning. This approach is especially relevant for probabilistic analysis of rare events, such as structural failure in safety critical systems, where not enough data exist to rely on observations alone. Some of the key challenges today relate to the rapid increase in structural complexity of engineering systems, together with the need for more accurate and reliable models to support decision making under uncertainty. Here, today's purely statistical approach is not sufficient. By introducing constraints based on phenomenological knowledge we believe that we can overcome these challenges, and develop the mathematical tools needed for probabilistic analysis of complex engineering systems in the digital era. It will thus form a basis for update of existing and development of future rules and recommended practices delivered by DNV GL. The main focus of the project is to develop new mathematical tools. To demonstrate how these tools may be applied, some relevant applications from the Oil & Gas industry will be selected for numerical experimentation.

Publications from Cristin

No publications found

No publications found

No publications found

No publications found

Funding scheme:

NAERINGSPH-Nærings-phd