Electrochemical neuromorphic devices, such as ionic synaptic transistors, offer promising opportunities for the development of energy-efficient, brain-inspired computing architectures. However, their design and optimisation remain challenging due to the complex coupling between ion transport, electrochemical processes and electrostatics. Current approaches largely rely on iterative fabrication and experimental characterisation, limiting the exploration of the design space and the identification of optimal device configurations.

This PhD project aims to develop a predictive, physics-informed digital-twin framework to accelerate the modelling and inverse design of electrochemical neuromorphic devices. The proposed approach combines multiphysics modelling, Bayesian inference and machine learning to predict device behaviour, quantify uncertainties and explore device architectures in silico, thereby reducing reliance on trial-and-error experimental approaches.

Starting from an existing Bayesian inference pipeline and neural-network-based surrogate model, the project will investigate advanced learning strategies to efficiently approximate the posterior probability distribution of physical and geometrical model parameters conditioned on experimental measurements and/or target device performances. Particular emphasis will be placed on physics-informed learning approaches combining governing PDE constraints with sparse, punctual experimental observations, inspired by Physics-Informed Neural Operators (PINO). Neural operators, conditional generative models (including diffusion models) and other state-of-the-art probabilistic learning strategies will be investigated to achieve accurate and computationally efficient posterior approximation, uncertainty quantification and inverse design.

Beyond electrochemical neuromorphic devices, the developed framework could be extended to a broader class of nonlinear dielectric systems, including ferroelectric and antiferroelectric devices. These systems involve strongly nonlinear constitutive relationships and complex coupled equations, often making conventional numerical modelling and parameter identification particularly challenging. They therefore represent a promising application for the proposed physics-informed Bayesian and machine-learning methodology.