3IA PhD/Postdoc Seminar #34

Flash presentation

A new high-performance computation method of the Eddington tensor in radiation hydrodynamics simulation

Abstract:
Radiative hydrodynamics models the coupling between the dynamics of a hypersonic hot plasma and the radiation it produces or external radiation. Almost every numerical code uses simplified models, which are often either limited or inaccurate. To accurately model photon transport, the HADES 2D code was specifically developed. Such a code is indispensable for studying astrophysical objects, in which optically intermediate regions are still poorly modeled but commonly encountered within such phenomena. This code couples hydrodynamics with the M1-multigroup model for radiation transfer to accurately represent the spectral behavior of light, involving the partitioning of the electromagnetic spectrum into groups. However, simulating radiative hydrodynamics flows remains highly time-consuming, constraining our capacity to conduct comprehensive numerical studies within this field.

The most computationally expensive part of the M1-multigroup simulations is the calculation of the closure relation, which relates the radiative pressure to the radiative energy and the radiative flux via the Eddington factor. This is due to the lack of an analytical solution. Consequently, two methods exist: One method is accurate but costly, relying on expensive search algorithms implemented in HADES. Another method is quicker but incorrect, utilizing the analytical grey case closure relation for each group, implemented in HERACLES.

To mitigate these challenges, we've pioneered an inventive approach intertwining neural networks with simplified models. This innovative method dramatically reduces computation time while maintaining an acceptable precision, revolutionizing the efficiency of these calculations within M1-multigroup simulations.

To affirm the efficiency of our approach, we conducted validation simulations, beginning with the renowned benchmark simulation of a 1D radiative shock, wherein we used up to five groups. Additionally, we undertook a radial test to assess the efficiency of our method in a 2D situation.

Deceptive AI and Society

Abstract:
‘Deceptive AI’ is an expression that captures the imagination of both users of AI technologies, and AI experts or researchers alike. But what does ‘Deceptive AI’ mean? Figuring out the meaning of this term is important to understand not just the recent hype around language models and their implementation in chatbot technologies, but also the historical gravitas behind building machines capable of deception, as well as how to take things forward from here. ‘Deceptive AI’ can refer to a multitude of different things, yet its entire range of meanings is intertwined with the human condition, i.e. how humans represent their own selves in the world, and the idea of hybrid societies, which are societies where humans and machines play the role of agents in a society.

In this presentation, I will take the audience through the various aspects that are relevant to deceptive AI research and try to explain the relations between the theories and technologies behind deceptive AI technologies, human-like intelligence, society, and the ethical principles and potential regulatory mechanisms that apply to deceptive AI technologies.

Open discussion about the two contributions