Research analysis · Neuromorphic hardware

A Hodgkin-Huxley neuron made of air valves

Researchers at KU Leuven and Cambridge have reformulated the Hodgkin-Huxley model of neuronal excitability as a domain-agnostic energy network, then mapped it onto pneumatic hardware: chambers for membrane capacitance, constant pressure manifolds for ionic reversal potentials, and pressure-controlled valves for voltage-gated ion channels. In simulation, the resulting fluidic neuron fires all-or-nothing pressure spikes and drives a half-center oscillator, with every branch dissipating non-negative power throughout.

Source: On Energy-Based Modelling for Fluidic Excitable Neurons, arXiv preprint 2611.16389, 12 Nov 2026. Primary source. Read the full 6-page preprint PDF.

What the work claims

This is a methods-and-simulation paper. The claim has two layers. First, a modeling claim: the Hodgkin-Huxley circuit, properly viewed, is built from only three classes of physical elements with distinct energetic roles - energy storage (the membrane capacitor), energy supply (the reversal potentials), and state-dependent dissipation (the ionic conductances, which are memductors, dissipative elements with memory).1 Because bond graph theory defines all three purely through power-conjugate effort and flow variables, the same topology can be translated into any physical domain that offers the right elements.

Second, a design claim: mapping that topology onto the fluidic domain - effort becomes gauge pressure, flow becomes volumetric flow rate - yields a physically realizable excitable neuron made of rigid gas chambers, restrictor tubes, and diaphragm-actuated slider valves, and the mapping is good enough that lumped pneumatic simulations produce genuine excitable behavior: threshold-triggered all-or-nothing spikes, sub-threshold relaxation, and rhythmic alternation in a two-neuron half-center oscillator driven by post-inhibitory rebound.1

The authors are explicit about what exists today: simulation only. No fluidic neuron has been built; the paper is a blueprint with numerical proof of concept.

How it works

The Hodgkin-Huxley equivalent circuit is five parallel branches across the membrane voltage: capacitance, external current, passive leak, sodium, and potassium. The authors re-express it as a bond graph: storage element C for the membrane capacitance; constant effort sources Se standing in for the three reversal potentials, which represent the electrochemical gradients that pay for the spike; and modulated resistors for the two voltage-gated conductances, whose state-dependence captures gating without modeling the underlying channel kinetics.1 The treatment follows the insight that the Hodgkin-Huxley axon is, energetically, a network of memristors.2

For hardware tractability they switch to a reduced two-state model from the mixed-feedback literature: a fast excitatory conductance driven by the membrane variable and a slow inhibitory conductance driven by a filtered recovery variable, each shaped by a static logistic sigmoid.3 The recovery variable is implemented as a passive RC low-pass filter, not an abstract integrated state. The energy balance is transparent: the storage function decreases through non-negative dissipation terms plus exchange with the input and the reversal supplies, so excitability arises from state-dependent access to the effort supplies rather than from any active power generation by the resistive elements, and no unphysical state reset is needed to terminate a spike.

The fluidic translation is one-to-one. The membrane capacitor becomes a rigid chamber whose compliance comes from gas compressibility (1 mL internal volume). The reversal potentials become regulated pressure lines: inhibitory exhaust at 0 bar gauge, excitatory supply at 1.0 bar gauge, leak biased at 0.2 bar gauge. The gated conductances become three-terminal ModulatedRestrictor valves in which a control pressure deflects a diaphragm that slides a profiled occluder across the flow channel. Branch parameters are concrete: 1.0 m reference tube length, 1.0 mm inner diameter for the leak path, 0.3 mm for the slow recovery path, yielding a membrane time constant of about 7 ms and a recovery time constant of about 0.9 s, a ratio near 10^-2 that produces sharp spikes with slow recovery.1

In the FONSim simulation environment for nonlinear inflatable actuator networks,5 a 50 ms mass-flow pulse of 12 mg/s elicits a single all-or-nothing pressure spike while a 10 mg/s pulse of the same duration does not; the membrane pressure simply relaxes. Instantaneous branch power remains non-negative at all times, which is the paper's central consistency check. Synapses are further modulated restrictors gated by the presynaptic recovery pressure, with a chosen reversal pressure setting excitation or inhibition, and two neurons cross-coupled by mutual inhibitory synapses produce sustained alternating spikes from a brief asymmetric perturbation, the classic half-center oscillator that vertebrate central pattern generators use for rhythmic locomotion, and a step beyond the limit-cycle fluidic oscillators previously demonstrated in soft robots, which generate rhythm but lack threshold-triggered excitability.41

Where a skeptic should push

The load-bearing assumption is that the idealizations survive contact with hardware. The gating modulation is modeled as an unloaded information signal: the control port that opens and closes the valve costs no energy and adds no dynamics. A real diaphragm-slider valve has mass, hysteresis, delay, and a power cost for switching, none of which appear in the model. The quasi-steady lumped pneumatic approximation neglects inertance and thermal dynamics, and the parameterization (a 1 m tube feeding a 1 mL chamber) is a bench-scale thought experiment, not a microfabricated device.

Also note the two-timescale structure that makes the spikes sharp: the 7 ms membrane constant against the 0.9 s recovery. That 100-fold separation is what the mixed-feedback motif needs, and in simulation it is guaranteed by choosing tube diameters. In physical valves, manufacturing tolerance, gas leakage, and compressibility variation with temperature all press on exactly this separation. And because every result is numerical, there is no noise analysis: real pneumatic systems are exquisitely sensitive to slow leaks, which the constant-pressure supply manifolds would have to fight continuously. None of this is fatal to the blueprint; all of it is unverified.

The substrate-agnostic physics of excitability

The non-obvious implication for biological computing is what the paper takes away and what it hands over. It removes the idea that excitability belongs to neurons. What makes a neuron a neuron, energetically, is a storage element, maintained supply gradients, and state-dependent dissipation; that recipe runs in electrochemistry, and the same topology runs in compressed air.1 For a field whose premise is that living neural tissue is a uniquely valuable computing substrate, this is a useful deflation: much of what makes tissue "computational" is generic excitable-media physics plus evolutionary tuning, not a monopoly of the living. The honest version of the organoid-intelligence thesis has to point at what tissue has that valves lack: self-repair, developmental scaling, synaptic specificity, metabolic self-maintenance of those supply gradients.

What the paper hands over is an accounting discipline. Biological computing discussions routinely treat spikes as nearly free because they are "event-driven." The bond-graph formulation makes the energetics explicit: every spike is paid for out of maintained gradients, and the model resolves instantaneous dissipation per branch, giving a direct metric for energy per spike that transfers conceptually to tissue, where metabolic cost accounting is usually hand-waved.1 An organoid-computing benchmark that reported dissipation-resolved energy per operation, in this framework's spirit, would be more honest than spike-count-per-watt comparisons against silicon.

The opportunity is hybrid interfacing. The authors' own motivation is timescale matching: electrical excitability runs at milliseconds while soft bodies move slowly, so pneumatic excitability, with its sub-second recovery, meets mechanical motion on its own clock.1 Organoid activity is slower still, often seconds per burst cycle. A fluidic intermediate layer whose natural timescales sit between electronics and tissue is a conceivable mediation technology: electronics read the dish, fluidics slow the signal into the tissue's regime, and the tissue's slow rhythms become legible events. The symmetric threat is competition: if event-based control primitives can be built from valves and chambers with no cells, no incubator, no ethics review, and no batch-to-batch biological variability, then the "embodied event-driven controller" niche that some biological-computing roadmaps claim is partly occupied by cheaper non-living matter. The half-center oscillator in this paper needs no feeding schedule.

The bottom line

Established: an energy-based reformulation of Hodgkin-Huxley that is domain-agnostic, and a pneumatic instantiation that is energetically consistent in simulation, producing genuine all-or-nothing excitability and rebound-driven rhythmicity. Not established: any physical device; all quantitative results come from a lumped, isothermal simulation with idealized unloaded control ports. What would confirm the blueprint: a fabricated valve network showing a measured all-or-nothing pressure spike and measured post-inhibitory rebound, with branch energetics matching the model. What would break it: valve dynamics or control-port loading that erodes the two-timescale separation, or supply systems unable to hold the reversal pressures against leakage over operating timescales.

Frequently asked questions

Is the fluidic neuron physically built?

No. Everything reported is simulation in the FONSim pneumatic network package. The paper is explicit that physical implementation is future work, so all numbers, including the 7 ms membrane time constant and the 0.9 s recovery, are model outputs.

What is a bond graph?

A domain-agnostic graphical formalism for modeling energy flow in physical systems. Components exchange power through bonds carrying effort and flow variables whose product is power; storage, supply, and dissipation elements are defined purely through those variables, which is what allows an electrical circuit to be translated into a pneumatic one.

What plays the role of ion channels?

Pressure-controlled valves called modulated restrictors. A control pressure deflects a diaphragm that moves a profiled slider across the flow channel, changing hydraulic resistance the way voltage gating changes ionic conductance. The reversal potentials become constant-pressure supply manifolds.

How fast is the simulated neuron?

The membrane time constant is about 7 ms and the slow recovery variable about 0.9 s, a separation of roughly 100-fold. Spikes are triggered by 50 ms input pulses: 12 mg/s is supra-threshold, 10 mg/s is not, in the reported parameterization.

Why does energetic consistency matter?

Because integrate-and-fire models terminate spikes with an abstract state reset that no physical device can perform for free. This construction derives spikes from storage, supply, and dissipation alone, with non-negative power in every branch at all times, so if built it would actually work energetically, and the model yields a per-spike energy cost rather than hiding it.

Could fluidic neurons couple to living tissue?

Nothing in the paper addresses biology, but the timescale logic cuts that way: pneumatic excitability runs at sub-second scales between silicon microseconds and organoid seconds. A fluidic layer is at least a conceivable slow, soft interface modality, though pressure coupling to tissue is unproven and the paper makes no such claim.

References

  1. R. Wroblowski, B. Michiels, B. Gorissen, and R. Sepulchre. On Energy-Based Modelling for Fluidic Excitable Neurons. arXiv preprint 2611.16389. 2026. https://arxiv.org/abs/2611.16389. Accessed 2026-11-14.
  2. L. Chua, V. Sbitnev, and H. Kim. Hodgkin-Huxley axon is made of memristors. International Journal of Bifurcation and Chaos 22(03), 1230011. 2012. https://doi.org/10.1142/S0218127412350011. Accessed 2026-11-14.
  3. R. Sepulchre, G. Drion, and A. Franci. Control across scales by positive and negative feedback. Annual Review of Control, Robotics, and Autonomous Systems 2, 89-113. 2019. https://doi.org/10.1146/annurev-control-053018-023708. Accessed 2026-11-14.
  4. L. C. Van Laake, J. De Vries, S. Malek Kani, and J. T. Overvelde. A fluidic relaxation oscillator for reprogrammable sequential actuation in soft robots. Matter 5(9), 2898-2917. 2022. Accessed 2026-11-14.
  5. A. Baeyens, B. Van Raemdonck, E. Milana, D. Reynaerts, and B. Gorissen. FONS: A Python framework for simulating nonlinear inflatable actuator networks. IEEE International Conference on Soft Robotics (RoboSoft), 1-6. 2023. Accessed 2026-11-14.