- 11 Feb 2025
-
-
Dept of Comp Sci Seminar: Dr Matt Ellis (University of Sheffield)
11 Feb 2025 2:00 pm - 3:00 pm
Kilburn_IT407Title: Noise aware training of spintronics-based physical neural networks
Abstract: Neuromorphic computing aims to reduce the energy footprint of AI technologies by mapping efficient biologically inspired structures onto physical hardware. Spintronic nano-devices have huge potential in this space due to their complex non-linear dynamics and non-volatility with efficient control methods. However, in many paradigms stochasticity and noise can be inhibitive, limiting off chip training. In this talk I will introduce two concepts for constructing physical neural networks from spintronic nano-devices and how noise can be incorporated into their design or training. The first is a racetrack neural network where magnetic domain walls are used as bits and probabilistic pinning at artificial notches are used to control their progress through the network in the same way as a synaptic weight. In this way the stochasticity of domain wall motion becomes a functional property and I will explain how repeated sampling can be balanced across training and test time. Secondly I will address training dynamical physical neural networks of multiple interconnected nano-devices. By training data driven stochastic models of the devices, the noise characteristics can be captured and used for off device training. This noise-aware training allows for better transfer to the hardware system, outperforming versions trained on deterministic models, when performed on dynamical tasks such as a smart prosthetic interface. Both these approaches demonstrate how noisy spintronic systems can be exploited as elements for unconventional computing.
BioDr Matthew Ellis is a Lecturer in Machine Learning and member of the Machine Learning Group at the Department of Computer Science at the University of Sheffield.
After completing his PhD in 2015 he joined the group of Prof. Stefano Sanvito at Trinity College Dublin as a post-doctoral research fellow. In 2019, he joined the University of Sheffield as a post-doctoral research associate in the Bio-Inpsired Machine Learning group under Prof. Eleni Vasilaki developing machine learning models for neuromorphic computing in collaboration with the Department of Materials Science.
Dr Ellis is interested in developing energy efficient machine learning algorithms and systems based on neuromorphic computing. In particular, he is interested in developing models of physical systems that can be utilised as machine learning processing devices, such as devices for physical reservoir computing or neuromorphic hardware based on magnetic systems. Beyond machine learning he is interested in developing large scale models of magnetic devices including developing gpu accelerated models.
-
- 12 Feb 2025
-
-
Mara Strungaru (University of Manchester)
12 Feb 2025 4:00 pm - 5:00 pm
Niels Bohr common room, 6.53Title:Advanced atomistic models of magnetic materials
Abstract: Magneticmaterials maintained huge interests for technological applicationssuch as magnetic recording media (initially on magnetic tapes, now onnm-sized granular media). With the increased necessity to store moreand more data, it is important to constantly improve/renew thecurrent technologies or seek for other magnetic entities (such asdomain walls, skyrmions) to act as a bit of information. Novelresearch fields in magnetism such as spin-electronics (spintronics),opto- and phono-magnetism, neuromorphic and reservoir computingpromise to bring more advanced technologies in our daily life, andthe usage of magnetic nano-particles in bio-medicine to even curesome types of cancer. The laser-induced manipulation of spinsalso promises to revolutionise the magnetic storage technologies byusing ultrafast processes with low dissipation.
Inthis work we will explore state-of-the art atomistic models ofmagnetic materials able to simulate billion atoms systems fortechnological applications and fundamental studies. I will start bypresenting the theoretical framework behind atomistic spin dynamics (ASD) which is based on solving the Landau-Lifshitz-Gilbert equationfor various magnetic interactions, implemented in an open-source codecalled VAMPIRE[1,2]. I will also present several applications of theatomistic model for recording media technologies (based on FePt) andnovel 2D magnetic materials, such as CrI3 [3]and CrCl3[4].I will then extend the atomistic spin dynamics to an unified model ofspin and molecular dynamics (spin-lattice dynamics -SLD) that takesinto account the lattice degrees of freedom (phonons). Such frameworkcanoffer a deeper understanding of magnon-phonon interactions [5,6],relaxation processes and phonon-driven switching mechanism, which canlead to the development of next-generation magnetic devices.
References
[1]R. F. Evans etal. Journal of Physics: Condensed Matter 26, 103202 (2014)
[3]M. Dabrowski et al Nature Communications 13, 5976 (2022)
[4]M.Strungaru et al npj Computational Materials 8, 169 (2022)
[5]Strungaru,Mara, et al. PhysicalReview B 109.22(2024)
[6]Strungaru, Mara, et al. PhysicalReview B 103.2(2021)
-
- 26 Feb 2025
-
-
Viktor Zolyomi (Daresbury Laboratory)
26 Feb 2025 4:00 pm - 5:00 pm
Niels Bohr common room, 6.53Title: "Computational Materials Discovery in the Era of Machine Learning and AI"
Abstract: "Materials discovery is in great demand across multiple sectors in industry due to drivers such as the Net Zero agenda. Adoption of new materials in industry faces challenges as the material space is vast, and trialing new formulations or metal alloy compositions in the lab is very expensive. The process can be both accelerated and de-risked by computational materials discovery, which can pre-screen candidate materials, potentially whittling down an initial longlist of 10000 materials to a shortlist of 10 in the time frame of 6-12 months. Computational modelling is however substantially expensive in its own right, often requiring first principles approaches. In recent years, machine learning and AI have undergone substantial advancements and we are now in a position to leverage these technologies to replace some critical steps in the computational modelling process with ML/AI tools that are several orders of magnitude faster yet deliver first principles accuracy. This talk explores how computational materials discovery accelerated by ML/AI technologies can be used for discovery of new materials or material compositions, through examples such as i) rapid discovery of organic compounds in search of active components for organic solar cells ii) pre-screening of high-entropy compound composition spaces to identify improved thermal barrier coatings needed in aerospace or the energy sector iii) developing rapidly deployable models for two-dimensional crystals with the ability to predict the performance of thin films in a range of applications from corrosion protection to catalysis control"
-