Physical Neuromorphic Computing

An alternative to the von Neumann paradigm

Digital processors keep memory and logic apart, which slows computation and raises energy use for data‑heavy tasks. Physical neuromorphic computing embeds the calculation in the material that stores the data. By harnessing intrinsic nonlinearities and memory effects, a physical neuromorphic device can perform complex transformations in a single physical step, often at a fraction of the energy required by conventional digital algorithms.

Neuromorphic reservoirs give unique advantages

In neuromorphic hardware the network’s connections and linked weights encode the algorithm. Training those weights usually dominates the efficiency of the system. Reservoir computers avoid this cost: their internal weights are fixed, and a lightweight external learner extracts the desired output from the reservoir’s high‑dimensional dynamical response. The reservoir thus acts as a physical kernel that expands an input into rich temporal and spatial features, while the read‑out merely fits a simple linear model.

All‑optical reservoirs and the promise of hybrid perovskites

Our group pursues two complementary reservoir approaches. First, we exploit hybrid metal‑halide perovskites, whose photoexcited state dynamics show opto-ionic behavior which can be read by probing their luminescence. The resulting electro‑ionic landscape evolves on nanosecond‑to‑microsecond scales, providing a natural high‑dimensional reservoir. By shaping an optical pump we encode the signal in the perovskite’s internal photoexcited states and let its intrinsic relaxation perform the reservoir computation.

Why this direction matters

Light‑based and perovskite reservoirs address three key challenges. They operate at speeds dictated by photon propagation and intrinsic material relaxation, outpacing conventional clock rates for many tasks. Their energy demand is limited to the optical pump, potentially at lower power than needed for digital weight updates in deep networks. Moreover, the same materials are compatible with established semiconductor and photonic fabrication lines, offering a realistic route to scalable neuromorphic hardware.

Beyond efficiency, studying these coupled electronic‑ionic processes deepens our understanding of phenomena that also govern perovskite solar‑cell stability and light‑emitting performance. Insights from the computational experiments therefore feed back into the broader optoelectronics community, creating an interdisciplinary feedback loop.