Developing data-driven workflows that merge micromagnetic simulations, optimization algorithms, and experimental feedback to design spin-wave components beyond the limits of intuition. We demonstrate how AI-guided design accelerates discovery and validation of complex magnonic devices.
Our machine learning programme targets the inverse design of spintronic and magnonic devices. We combine differentiable simulations, surrogate modelling, and reinforcement learning to identify geometries that realize desired dispersion relations and functional responses.
Close coupling between computation and fabrication enables rapid iteration. Automated experimental validation feeds back into our models, continually refining predictions and reducing reliance on brute-force parameter sweeps.
This work culminated in the first experimentally verified spin-wave lens conceived through machine learning, establishing a blueprint for AI-assisted discovery in nanomagnetism.
Realized a high-efficiency magnonic lens optimized entirely through machine learning and validated on focused-ion-beam patterned YIG films.
Deployed a closed experimental loop that updates neural network surrogates with each fabrication run, cutting optimization time by an order of magnitude.
Embedded conservation laws and dispersion constraints into neural architectures, enabling generalizable predictions across frequency bands and materials.
Released modular Python toolkits that integrate micromagnetic solvers, differentiable programming, and laboratory automation scripts for community reuse.
Run large-scale micromagnetic simulations and targeted experiments to populate training sets spanning geometric and material variations.
Employ Bayesian optimization, differentiable programming, and reinforcement learning to search high-dimensional design spaces efficiently.
Translate digital designs into devices using FIB patterning and lithography, with BLS and TR-MOKE measurements closing the loop for continued learning.
Combine expert heuristics with model-driven suggestions to maintain interpretability and accelerate iteration cycles.
Optimize for tolerance against fabrication variability and thermal noise, ensuring real-world reliability of AI-generated geometries.
Integrate laboratory automation workflows that allow thousands of virtual experiments per night and rapid experimental follow-up the next morning.
First experimental realization of a spin-wave lens designed using machine learning optimization techniques.
Extend diffusion and transformer-based models to synthesize candidate device layouts that meet multiple objectives simultaneously.
Incorporate live measurements into training loops to adapt device controllers and compensate for drift in real time.
Couple magnonic, photonic, and electronic simulations within a unified optimisation framework to co-design hybrid platforms.
Grow community-driven tools and datasets that lower barriers for AI-assisted device design in magnetics and beyond.