Back to search

FRINATEK-Fri prosj.st. mat.,naturv.,tek

Deep learning the real-time properties of strongly correlated quantum fields

Alternative title: Dyplæring anvendt på beregning av dynamikken til sterkt korrelerte kvantefelt

Awarded: NOK 7.0 mill.

20th century physics revolutionized our understanding of the matter around us on microscopic scales. Inside protons and neutrons we find three quarks, held together by elementary particles aptly referred to as glue. But our universe has not always been as calm as it is today. Shortly after the BigBang temperatures were 200000 times hotter than the center of our sun. In the 21st century high energy physics strives to understand how quarks and gluons behave under such extreme conditions. To this end large accelerators have been built, in which heavy ions are collided, recreating conditions similar to the BigBang. Describing matter in extreme conditions requires quantum field theory. It represents a theoretical framework, which captures the fact that particles can be created from energy via E=mc^2. There is strong evidence that the field theory Quantum-Chromo-Dynamics (QCD) correctly describes quarks and gluons. But even after decades of work it has not been possible to accurately compute their real-time dynamics under the conditions in heavy-ion colliders. The reason is that quarks and gluons are strongly coupled and all known weak-coupling methods fail. On the other hand, numerical simulations of QCD suffer from the so called sign-problem, making them unstable. DeepRTP develops new strategies to compute the dynamical evolution in QCD under extreme conditions. The novelty of our approach is to utilize recent advances in the field of machine learning. We propose to both improve existing computations methods, i.e. minimize the influence of the sign problem, while at the same time develop genuinely novel simulations attempting to avoid the sign problem all together. Gaining access to the dynamics of strongly correlated quantum fields will not only significantly improve our understanding of the early universe but will also open novel paths of research in other fields, e.g. the study of ultracold quantum gases under extreme conditions and quantum information science. Recent research highlights arising from this project include the first determination of the interaction potential between heavy subatomic quark particles in a real-time simulation and a novel implementation of a promising real-time simulation approach called Complex Langevin, significantly improving its stability. In 2022, we reached another milestone. We used machine learning techniques to extend the range of Complex Langevin simulations of the dynamics of quantum fields to three times the previous record published in the literature, presented at the international symposium of Lattice field theory in Bonn, Germany. In 2023 we made a key breakthrough by being able to simulate a more realistic system, so called scalar field theory in two dimensions, up to twice the time extent previously reported in the literature, establishing our method as the current best approach to the simulation of real-time dynamics of this class of quantum systems. Concurrently the project is progressing towards its goal of extracting the dynamics of subatomic quark particles from conventional simulations, playing a vital role in an international collaborative effort on the subject. For current project information see our homepage at http://deeprtp.uis.no

The ab-initio study of quantum systems of more than 10 particles and the real-time quantum evolution of fields that describe how light and matter interact is severely limited by the so-called sign-problem. This problem hampers progress in high energy physics, material science and quantum chemistry. The DeepRTP project (w/ PhD: D. Alvestad, Postdoc: R. Larsen) developed a novel and innovative approach to tackle the sign problem by using machine learning. Combining modern stable simulation algorithms developed for financial mathematics applications with reinforcement learning strategies, proven successful in autonomous driving, we were able to infuse simulations of quantum fields with additional relevant knowledge. This allowed us to stabilise and extend the range of validity of these simulations significantly beyond the state-of-the-art, beating several records previously set in the literature. With these impactful contributions, DeepRTP has propelled ahead and reinvigorated the study of real-time quantum dynamics with the complex-Langevin approach, seeing related projects take shape in Austria, Hungary and the Denmark. As second pillar of DeepRTP we deployed advanced data analysis techniques to extract for the first time the potential between a heavy quark and antiquark at finite temperature from state-of-the-art numerical simulations of the strong nuclear force (w/ PhD: G. Parkar, Postdoc: R. Larsen). Our results upset the established picture of heavy quark interactions, as it revealed that contrary to expectations, the potential does not appear to be screened in the presence of mobile color charge carriers. Such a result would have far reaching consequences for the dynamic modelling of heavy quark pairs relevant for ongoing and upcoming heavy-collisions experiments carried out at the CERN and BNL laboratories. As part of this DeepRTP pillar, we co-developed a deep-learning based strategy for the extraction of so-called spectral functions, taking on a pioneer role in this field together with colleagues from the University of Heidelberg.

The DeepRT project combines methods from theoretical physics and machine learning to tackle a central physics challenge that hampers progress in many areas, e.g. relativistic heavy-ion collisions. Our physics goal is to elucidate how energy and momentum are transferred among microscopic matter constituents in extreme conditions, where "extreme" may refer to high temperatures or density. It requires us to compute the transport coefficients bulk and shear viscosity. In quantum field theory, such real-time properties are encoded in spectral functions. Extracting those from standard first-principles Monte-Carlo simulations however amounts to an exponentially hard inverse problem. We aim to overcome this issue with an improved extraction (subproject I) and an alternative simulation prescription (subproject II). In subproject I a PhD student will apply the concept of autoencoder (AE), a pillar of the deep-learning revolution in image processing, to the extraction of spectral functions from Monte-Carlo simulations. The central task is to develop the required topology and training strategy for the neural network. A large data pool from previous work of the PI will be an essential training ingredient. The ability of an AE to represent relevant features of input in hidden layers will be used to obtain improved Bayesian regularization schemes for inverse problems. In subproject II a postdoc will combine deep learning with complex Langevin simulations to compute real-time properties directly, evading the inverse problem. The Langevin approach so far has suffered from instabilities, which we propose to overcome by a neural network that acts as a dynamical control system. Developing and training the network in both the non-interacting theory, where analytic data is available and from concurrent standard imaginary time simulations will be the central challenge. In case of gauge fields, the degeneracy in the field variables will require developing and handling of deep networks.

Publications from Cristin

No publications found

No publications found

Funding scheme:

FRINATEK-Fri prosj.st. mat.,naturv.,tek