Link to the code: brain-emulation GitHub repository

Neuromorphic Twins: Digital Brain Models That Grow With You

Most digital brain models are static snapshots. They capture the state of a neural network at a moment in time — its connectivity, its firing rates, its response properties — and simulate that state. They do not change when the biological brain changes. If the patient undergoes learning, injury, or disease progression, the model becomes outdated. It no longer represents the brain it was built to model.

A 2026 paper in Nature Communications describes a different approach. Neuromorphic Twin technology creates digital models of biological neuronal networks that run in real-time and co-evolve with the living tissue they represent. As the biological brain learns, adapts, or degrades, the digital twin tracks those changes. The two systems remain synchronized — not as a snapshot and its record, but as two dynamically coupled representations of the same neural substrate.

How Neuromorphic Twins Work

The Neuromorphic Twin architecture has three components: a biological recording interface, a neuromorphic computing substrate, and a bidirectional coupling protocol.

The biological interface captures neural activity through high-density electrode arrays, optogenetic readouts, or multielectrode recordings. This data stream represents the current state of the biological network: which neurons fired, in what sequence, with what interspike intervals. The interface does not need to capture every neuron — statistical sampling and computational inference fill gaps based on population dynamics.

The neuromorphic substrate runs a model of the network on spike-based hardware — silicon implementations of spiking neural networks (SNNs) that operate in the same temporal domain as biological neurons. Unlike rate-coded neural network models (which compress time and use floating-point activations), neuromorphic implementations fire discrete spikes with millisecond-level timing. This temporal fidelity is what allows real-time synchronization with biological tissue.

The coupling protocol is what distinguishes the Neuromorphic Twin from prior digital twin work. It is bidirectional: biological activity updates the digital model, and the digital model’s predictions can be used to generate stimulation patterns fed back to the biological tissue. The biological and digital networks do not operate in isolation. They influence each other.

What “Co-Evolution” Means in Practice

Co-evolution in this context means that plasticity in the biological network is mirrored in the digital twin, and vice versa. When biological synapses strengthen through repeated activation — a Hebbian learning process — the corresponding synaptic weights in the neuromorphic model are updated to match. When the biological network weakens connections through long-term depression, the digital twin degrades the same connections.

The model’s plasticity rules are tuned to match the specific learning dynamics of the biological tissue being modeled. This is not a generic SNN model running a standard plasticity algorithm — it is a personalized model calibrated to the individual’s neural learning rates, synaptic time constants, and network topology.

The authors demonstrated this bidirectional co-evolution in vitro using dissociated cortical neuron cultures and in vivo using rodent hippocampal circuits. In both cases, the digital twin tracked learned changes in the biological network with high fidelity over periods of days to weeks. When the researchers perturbed the biological network — through pharmacological intervention or targeted stimulation — the digital twin captured the resulting changes and could predict network state after the perturbation.

The Connection to Whole Brain Emulation

The Virtual Brain Twins research covered earlier on this blog described personalized brain models built from structural imaging data — individualized simulations of whole-brain dynamics for clinical use. Those models are powerful but static: they represent the brain at the time of imaging and do not update as the biological brain changes.

The Neuromorphic Twin concept is complementary and, in some ways, more radical. It is not built from imaging data. It is built from ongoing biological activity and maintained through continuous synchronization. It does not represent what a brain looked like at scan time. It represents what the brain is doing now — and it has been doing so continuously since calibration.

This distinction matters for whole brain emulation in two ways.

First, a Neuromorphic Twin is intrinsically temporal. WBE has long wrestled with the problem of when to take the snapshot that becomes the emulation. A brain that is scanned destructively (as in the Song et al. preservation protocol) is captured at one moment. The emulated mind is frozen at that moment’s state, unable to incorporate any learning or experience from between scan and instantiation. A continuously synchronized digital twin has no this problem — it is always current.

Second, the bidirectional coupling enables a gradual transition scenario. If a biological brain is continuously coupled to a neuromorphic twin, and the coupling can be gradually shifted so that more and more cognitive function is handled by the digital system, then the transition from biological to digital cognition could be incremental rather than abrupt. This connects directly to Hans Moravec’s 1988 gradual neuron replacement proposal, but implemented at the network level rather than the neuron-by-neuron level.

The paper’s authors do not frame their work in terms of mind uploading. The immediate applications they discuss are medical: personalized brain disorder treatment, real-time neural prosthetics, closed-loop neuromodulation for epilepsy, Parkinson’s disease, and depression. But the technical architecture they describe — a continuously synchronized, co-evolving digital replica of a living brain — is directly relevant to the WBE question.

Personalized Treatment and the Path Forward

The clinical applications are compelling on their own terms. Current deep brain stimulation (DBS) for Parkinson’s disease uses fixed stimulation parameters that are set at implant time and adjusted infrequently in clinical follow-ups. The stimulation does not adapt to moment-to-moment changes in the patient’s neural state. A Neuromorphic Twin-based closed-loop system would use the digital model’s predictions to adjust stimulation parameters continuously, providing personalized, adaptive neuromodulation in real time.

The same principle applies to epilepsy. Seizure prediction currently relies on detecting the early signatures of seizure onset in raw neural signals. A digital twin that models the patient’s specific network dynamics could predict seizures earlier, detect subtle precursor states invisible in raw signals, and trigger preventive stimulation before seizure onset.

For depression and treatment-resistant psychiatric conditions, the Neuromorphic Twin opens the possibility of testing stimulation protocols in the digital model before applying them to biological tissue. The ability to run virtual experiments on a patient’s personalized digital twin reduces the risk and time cost of optimizing neuromodulation therapy.

Technical Constraints

The current implementation has significant limitations. The largest Neuromorphic Twin demonstrated in the paper covered a network of several thousand neurons — a small fraction of even a local cortical column, which contains tens of thousands of neurons. Scaling to the millions of neurons in a cortical region, let alone the billions in a complete brain, requires neuromorphic hardware that does not yet exist at the necessary density and interconnect bandwidth.

The coupling protocol requires high-density neural recording, which is currently limited to either invasive electrode arrays or lower-resolution noninvasive methods. A whole-brain Neuromorphic Twin would require either a completely solved problem in noninvasive neural recording or a large-scale implanted interface — neither of which is near-term.

The real-time synchronization requirement is also computationally demanding. Biological neurons operate at millisecond timescales. A real-time digital twin must process incoming neural data and update model state within that same window, continuously, without lag. Current neuromorphic hardware (Intel Loihi 3, IBM NorthPole) approaches this requirement at small scales but would require substantial architectural development for brain-scale operation.

These are engineering challenges, not theoretical barriers. The Neuromorphic Twin concept demonstrates that the bidirectional, co-evolving architecture works. Scaling it is a defined problem, not an undefined one.

Official Sources