AI-driven Simulations for Neuromorphic Spintronics
| Title: |
AI-driven Simulations for Neuromorphic Spintronics |
| DNr: |
Berzelius-2026-335 |
| Project Type: |
LiU Berzelius |
| Principal Investigator: |
Alexander Edström <aleeds@kth.se> |
| Affiliation: |
Kungliga Tekniska högskolan |
| Duration: |
2026-11-12 – 2027-06-01 |
| Classification: |
10304 |
| Homepage: |
https://www.kth.se/profile/aleeds?l=en |
| Keywords: |
|
Abstract
Magnetic materials host a remarkable variety of phases: collinear and noncollinear antiferromagnets, spin spirals, skyrmion lattices, frustrated and glassy states. These phases underpin technologies ranging from data storage and spintronics to energy-efficient and neuromorphic computing. Finding out which phase a material adopts, and how it responds to strain, fields or chemical substitution, means exploring rugged, high-dimensional energy landscapes. This remains costly even with modern first-principles and atomistic spin-dynamics methods, and labelled reference data for training conventional supervised models are scarce.
This project is part of an interdisciplinary WISE–WASP NEST collaboration between KTH, Uppsala University and Örebro University. It develops and applies brain-inspired artificial intelligence, specifically the Bayesian Confidence Propagation Neural Network (BCPNN), to atomistic spin models. BCPNN is a cortex-inspired network built from modular hypercolumn/minicolumn units with local Bayesian–Hebbian learning, sparse distributed representations and attractor (associative-memory) dynamics. It learns efficiently from unlabelled data in unsupervised and semi-supervised settings, which makes it well suited to materials problems where labels are expensive or ambiguous.
We will train BCPNN on large sets of spin configurations generated by atomistic spin-dynamics and Monte Carlo simulations of realistic spin Hamiltonians parametrised from density functional theory, so that it learns hidden representations that identify magnetic phases and phase transitions without prior labels. We will also use BCPNN’s associative-memory dynamics to navigate the complex energy landscapes of spin models, recalling low-energy and metastable configurations from partial or noisy inputs.
The goal is to establish brain-inspired AI as an efficient, data-frugal tool for discovering magnetic phases. We expect three kinds of impact. For materials science, fast and systematic identification of exotic magnetic phases will help guide experiments and materials design, while also potentially providing new physical insights. For AI for science, the project will show what brain-like learning offers on a demanding physical problem, beyond standard machine-learning benchmarks. Finally, it connects the two fields: magnetic systems whose collective dynamics resemble associative-memory computation are natural candidates for neuromorphic hardware, so the project also informs the design of magnetic materials with functionality for neuromorphic computing.