Multimodal training in reconfigurable brain-inspired iontronic networks

Inspired by the brain, we propose an iontronic physical alternative to traditional digital neural networks. We theoretically study microfluidic networks in which nodes are connected by conical electrolyte-filled channels that act as memristive iontronic synapses. Their electrical conductance is not only tunable via the channel geometry, but also depends on electrical, chemical, and mechanical signals. Leveraging this multimodal responsiveness, we demonstrate that small networks can be trained to perform nonlinear tasks. The network performs forward passes physically via ionic relaxation, while training combines this physical evolution with numerical gradient descent. We theoretically demonstrate that this system can perform tasks like input-output mapping and linear regression with bias. Moreover, we show that the circuit can be reconfigured to perform different objectives via adjusting the pressures on the nodes, paving the way for soft adaptable materials that compute and learn without conventional electronics.