DQMP Forum - Ferroelectricity and Superconductivity in 18O-substituted SrTiO3—delta - Scientific Machine Learning

16.06.2020 13:00 – 14:30

Ferroelectricity and Superconductivity in 18O-substituted SrTiO3—delta
Dorota Pulmannova (group of Prof. van der Marel)

SrTiO3 is an insulator which becomes superconducting upon electron doping at an exceptionally low carrier density, making it the most dilute superconductor. In the normal state, SrTiO3 is a quantum paraelectric with a remarkably high dielectric constant. Quantum fluctuations can be suppressed by 18O substitution, driving the system into a ferroelectric state. The superconducting and ferroelectric orders may be accidental neighbors or intimately connected, as in the theory of quantum critical ferroelectricity. The mechanism of the superconducting pairing, however, is still a matter of debate.
We have developed a process to substitute 18O for 16O in strontium titanate single crystals, up to a substitution level of 80 atomic %, as determined from the increase of the mass of the sample. Dielectric measurements confirm the ferroelectric order at low temperatures and homogeneous distribution of 18O in the samples.
Beyond the critical threshold of 33 % 18O, the superconducting critical temperature of the oxygen deficient SrTiO3−δ crystals is strongly enhanced. This supports the role of the vicinity of the ferroelectric order in the superconducting pairing.


Scientific Machine Learning
Charles Bardyn (group of prof. Giamarchi)

Machine learning techniques based on differentiable programming have led to impressive advances such as self-driving cars and StarCraft II Grandmaster-level bots. For the most part, these systems were programmed automatically using a large amount of data instead of hard-coded rules such as physics laws. In contrast, scientific models have traditionally been developed using human-driven modelling alone, with data mostly restricted to model testing. In this talk, I discuss how the two scientific approaches can be combined into scientific or science-informed machine learning, where the role of experimental data is elevated from parameter fitting to full-fledge modelling. I show examples of how scientific models can be completed or reversed engineered automatically from data using differentiable programming and trainable universal function approximators such as neural networks.
I come in peace with one message: that computers, classical or quantum, are generally more expressible when programmed in a differentiable way and optimized automatically towards a targeted task.

Lieu

Zoom Meeting

Please join us on Zoom, Meeting ID: 582 067 708
https://unige.zoom.us/j/582067708

entrée libre