A new computational study proposes an unusual way to explore diabetic peripheral neuropathy: by using a simulated perovskite memdiode to reproduce how damaged sensory systems might lose the ability to learn, adapt, and predict. Published in Neuroinformatics, the work combines emerging neuromorphic hardware concepts with predictive-coding theory, creating an in silico surrogate model that links device physics to the computational consequences of neuropathy. The researchers, Deiva Kumar K. and Mathivanan Ponnambalam of Amrita Vishwa Vidyapeetham in India, present the model not as a medical diagnostic tool, but as a virtual laboratory for studying how multiple neurological failures could interact inside a predictive brain-like circuit.
The centerpiece of the study is the halide perovskite memdiode, a device whose electrical behavior is shaped by both electronic charge and the movement of ions. Unlike a conventional resistor, whose resistance is ideally fixed, a memdiode can retain a history of previous electrical stimulation. Its conductance changes according to the timing, strength, and duration of applied pulses, allowing it to imitate some of the adaptive properties of biological synapses. The researchers incorporated a comprehensive mathematical representation of this coupled ionic-electronic behavior into a continuous-time neuromorphic circuit. In practical terms, the circuit treats the simulated device as a dynamic synapse whose internal state evolves rather than simply switching between zero and one.
Before applying the device model to neuropathy, the researchers tested whether it could reproduce a broad repertoire of synaptic learning behaviors. Simulations generated analog long-term potentiation and long-term depression, the strengthening and weakening processes commonly associated with memory formation. The model also reproduced spike-timing-dependent plasticity, in which the order and time difference between pre- and postsynaptic spikes determine whether a connection strengthens or weakens. In addition, it captured spike-rate-dependent plasticity, paired-pulse facilitation, and paired-pulse depression. These tests are important because a neuromorphic component intended to represent a biological synapse must respond not only to individual pulses, but also to patterns unfolding across multiple timescales.
The simulated memdiode also displayed pulse-amplitude- and pulse-width-dependent conductance modulation. Stronger or longer electrical inputs produced different internal-state trajectories, reflecting the fact that ionic redistribution within a perovskite device can occur gradually and can persist after stimulation stops. The model generated pinched hysteresis loops in its current-voltage response, a hallmark often associated with memory-dependent circuit elements. Such behavior arises because the current at a given voltage depends partly on the device’s prior history. The researchers further used Monte Carlo simulations to introduce device-to-device and cycle-to-cycle fluctuations, testing whether the proposed synaptic responses survived realistic statistical variation rather than appearing only under perfectly controlled conditions.
The study then placed the memdiode inside a predictive-coding framework. Predictive coding describes perception as an ongoing exchange between internal predictions and incoming sensory evidence. A network attempts to reduce prediction errors, and the researchers used variational free energy, or VFE, as a computational measure of how well the system reconciled its predictions with its simulated sensory input. In the model, synaptic adaptation was not an isolated learning rule; it became part of a larger feedback system that continuously adjusted internal states. This allowed the researchers to examine how impaired plasticity, delayed signaling, weakened sensory input, and disrupted homeostasis could collectively interfere with the network’s ability to minimize VFE.
Four clinical hallmarks associated with diabetic peripheral neuropathy were translated into specific circuit parameters. Impaired plasticity was represented through the parameter (k{drift}), which controls the rate of internal conductance-state evolution. Homeostatic failure was linked to the threshold-related timescale (\tau{TH}), which influences how the system regulates its activity and learning threshold. Afferent attenuation was represented by (A{amp}), reducing the amplitude of incoming sensory information, while conduction delay was modeled using (\tau{delay}), introducing a temporal gap between a simulated stimulus and its arrival in the network. This mapping does not claim that these four parameters directly correspond to isolated biological mechanisms. Instead, it provides a transparent computational vocabulary for exploring how distinct forms of dysfunction might alter predictive processing.
To control the overall level of simulated disease, the researchers introduced a single severity parameter, α. At the healthy baseline, the predictive-coding network reduced its VFE rapidly, indicating that it could adjust its internal estimates to match incoming information. At α = 0.5, representing a moderate level of modeled dysfunction, VFE remained elevated and oscillatory rather than settling quickly. The authors describe this computational pattern as a hypothetical counterpart of allodynia, a condition in which normally harmless stimuli can become painful. The analogy is deliberately conceptual: the circuit does not feel pain, and the simulation does not reproduce the complete biological mechanisms of allodynia. It shows instead how unstable prediction-error dynamics could provide a theoretical bridge between altered sensory input and persistent abnormal inference.
At severe modeled neuropathy, with α values of 0.8 or higher, the network entered a more dramatic regime. Homeostatic plasticity became effectively paralyzed, leaving the circuit trapped in a frozen maladaptive state. Rather than continuing to update its predictions, the system lost the flexibility needed to correct persistent errors. The authors compare this behavior computationally with sensory ataxia, in which impaired sensory feedback can disrupt coordination and the perception of body position. Again, the comparison is not a clinical simulation of a patient. It is a proposed systems-level metaphor in which the failure to update internal representations leads to a rigid and unreliable model of the outside world.
The investigators also examined how noise and parameter changes affected the circuit. Their noise-resilience analysis suggested that the major dynamical regimes remained recognizable across the stochastic conditions tested, while hyperparameter sensitivity experiments indicated structural robustness to perturbations of model settings. This result is significant for neuromorphic research because real devices rarely behave identically from one cycle to the next. Perovskite materials can exhibit variability caused by ionic motion, defects, interfaces, fabrication differences, and environmental conditions. A model that collapses under tiny parameter changes would offer little guidance for hardware design. However, the robustness reported here applies only within the simulated parameter space. It does not establish that physical perovskite memdiodes will display the same stability, endurance, or reproducibility.
The broader significance of the work lies in its attempt to connect three traditionally separate areas: perovskite device physics, brain-inspired learning, and computational models of diabetic neuropathy. Earlier neuromorphic studies have shown that halide perovskites can support memory, synaptic behavior, and higher-order temporal dynamics. This study extends that direction by asking what such dynamics might mean when embedded in a predictive-coding architecture. Its most ambitious claim is not that a memdiode can diagnose neuropathy, but that second-order, BCM-capable device dynamics can serve as a building block for modeling how sensory systems fail. The publicly available repository includes LTSpice netlists, subcircuits, parameter tables, severity-dependent equations, Monte Carlo settings, and scripts intended to make the framework reproducible.
The researchers emphasize that the model is an illustrative computational surrogate, not a validated biological replica or a treatment platform. Its clinical hallmarks are translated into circuit parameters through theoretical assumptions, and the reported disease-like states emerge from simulations rather than recordings from patients or experiments on peripheral nerves. The next challenge will be to compare the model’s predictions with physiological and behavioral data, determine whether its parameter mapping can be independently calibrated, and establish whether real perovskite memdiodes can reproduce the required dynamics under practical operating conditions. Even with those limitations, the study offers a striking vision of neuromorphic science: a future in which the failure modes of artificial synapses are used not only to build intelligent machines, but also to explore how biological intelligence might unravel when learning, timing, sensation, and homeostasis begin to drift apart.
Subject of Research: Neuromorphic computing, perovskite memdiodes, predictive coding, and computational modeling of diabetic peripheral neuropathy
Article Title: A Perovskite Memdiode-Based Neuromorphic in Silico Surrogate Model for Emulating Predictive Coding Failure in Diabetic Neuropathy
Article References: K, D. K., & Ponnambalam, M. (2026). A Perovskite Memdiode-Based Neuromorphic in Silico Surrogate Model for Emulating Predictive Coding Failure in Diabetic Neuropathy. Neuroinformatics, 24, Article 55.
Image Credits: AI Generated
DOI: 10.1007/s12021-026-09809-x
Keywords: Neuromorphic computing; perovskite memdiode; diabetic peripheral neuropathy; predictive coding; variational free energy; synaptic plasticity; BCM learning rule
