Skip to main content

CSC@Lunch Seminar Series

We share a weekly seminar series with the Warwick Centre for Predictive Modelling (WCPM). Seminars are held from 1-2 pm on Mondays during the university term. Nominations for speakers are welcome. Please contact James Kermode or Peter Brommer with suggestions.

To receives updates and reminders about the series, please subscribe to the csc-events mailing list.

 
 
Mon 29 Apr, '19
-
Joseph Barker (Leeds)
PS0.17 Physical Sciences

Atomistic spin dynamics with a quantum thermostat

Atomistic spin dynamics is a common method used to calculate dynamics and thermodynamics of magnets. It is a classical formalism based on the Heisenberg Hamiltonian. Magnetic materials can be modelled in exquisite detail, with the exchange and additional Hamiltonian terms parameterised from ab initio or experiments. It is a good approach for modelling complex magnets where simple magnon band theories can be inadequate. Even though such complex models can be built, the formalism still lacks quantitative power because classical (Rayleigh-Jeans) statistics are generally used, which are inappropriate at low temperatures. This is equivalent to the ultraviolet catastrophe of black-body physics, but for magnons.

We have incorporated a quantum thermostat into atomistic spin dynamics so that the magnons now obey Planck statistics. This allows truly quantitative calculations to be performed. We apply this method to calculate thermodynamic and magnon transport properties in the complex ferrimagnet yttrium iron garnet (YIG). This magnetic insulator is used across many research fields due to its ultra-low Gilbert damping. The large unit cell, containing 20 magnetic atoms cannot be approximated easily, as we will show. We have calculated thermodynamic quantities of interest in spintronics—such as the magnon heat capacity and magnon spin conductivity. These are extremely difficult to measure in experiments and often limited to only the low temperature regime. Our calculations at low temperature show excellent agreement with experimental measurements. Calculating beyond this regime we show the deficiency of analytic methods in such complex systems, due to their crude approximation of the magnon spectrum.

The code I develop is now about 10 years old and began in the days of CUDA 0.7. Time permitting I will also give a brief description of how we use GPUs to accelerate our calculations.

Mon 13 May, '19
-
Geneviève Dusson (Warwick Mathematics)
D2.02 Engineering

Physics+data-driven interatomic potentials based on permutation-invariant polynomials

Mon 20 May, '19
-
David Parfitt (Coventry)
PS0.17 Physical Sciences

Atomic Scale Modelling of Diffusion in Energy Materials

Mon 3 Jun, '19
-
Henrik Singmann (Warwick Psychology)
D2.02 Engineering

A Bayesian and Frequentist Multiverse Pipeline for Multinomial Processing Tree Models – Applications to Recognition Memory

Authors: Henrik Singmann (University of Warwick), Daniel W. Heck (Universität Mannheim), Marius Barth (Universität zu Köln), Julia Groß (Heinrich-Heine-Universität Düsseldorf), Beatrice G. Kuhlmann (Universität Mannheim)

Abstract: Even with a clear hypothesis or cognitive model in mind, most statistical analyses contain several more or less arbitrary choices. In the case of a model-based analysis, these choices can concern the statistical framework, the aggregation-level, and which parameter restrictions to introduce. Usually one path through this ‘garden of forking paths’ (Gelman & Loken, 2013) is chosen and reported. However, it is unclear how much each choice affects the reported results. The multiverse approach (Steegen, Tuerlinckx, Gelman, & Vanpaemel, 2016) offers a principled alternative in which results for all possible combinations of reasonable modeling choices are reported. We developed a software package for R that performs a model-based multiverse analysis for multinomial processing tree (MPT) models, MPTmultiverse. Our package estimates MPT models in a frequentist and Bayesian manner. In the frequentist case, it uses no pooling (with and without bootstrap) and complete pooling. In the Bayesian case, it uses no pooling, complete pooling, and three different variants of partial pooling. We applied our approach to a large confidence-rating recognition memory data corpus consisting of 12 studies with over 450 participants using a relatively unrestricted variant of the 2-high threshold model for confidence ratings (Bröder, Kellen, Schütz, & Rohrmeier, 2013). Our results show that even for some core parameters, the different analysis approaches reveal considerable variability in the parameter estimates across estimation methods. Our results suggest that researchers should adopt a multiverse approach when using cognitive models.

Mon 10 Jun, '19
-
Sondipon Adhikari (Swansea)
PS0.17 Physical Sciences

TBC

Mon 17 Jun, '19
-
Joanna Leng (Leeds)
D2.02 Engineering

The Importance of Exploratory Visualization for Research

Mon 24 Jun, '19
-
Francesca Baletto (KCL)
PS0.17 Physical Sciences

TBC