Research analysis · Computing substrates

Computing with waves: photonics moves into organoid territory

A focused review from Georgia Tech and the National Hellenic Research Foundation lays out the mathematical machinery by which metamaterials and integrated photonic circuits perform matrix operations, solve equations and run neural-network inference directly in propagating waves. For anyone who argues that computation should be delegated to exotic physical substrates, this is the competition's playbook, and its honesty about measurement boundaries is stricter than anything the organoid field currently practices.

Source: Metadevices that compute: mathematical foundations and experimental implementations for neuromorphic nanophotonics, arXiv, 2026-10-29. Primary source. Read: full text, 37-page PDF.

What the work claims

This is a review and synthesis, not a primary experimental result, and it should be weighted accordingly: its strength is consolidation and framing, not new data. Hadia, Tsilipakos, Tzarouchis and Pestourie claim that a coherent mathematical stack now exists for wave-based analog computing. Engineered scattering processes implement linear operators; feedback architectures implement equation inversion; inverse design and scientific machine learning translate a desired computation into a physically fabricable structure; and a growing body of experiments, from free-space diffractive processors to reconfigurable guided-wave circuits, demonstrates the pieces across radio frequency, microwave, terahertz and optical regimes1.

Two framing choices make this review different from the usual photonic-computing survey. First, it is explicit that fixed linear operators and equation solvers are not themselves neuromorphic; they are primitives that become neuromorphic only when composed with feedback, nonlinearity and task-dependent learning. Second, its performance comparison chapter refuses to collapse numbers onto one axis: every platform's throughput and efficiency figure is annotated with its reporting basis, power boundary and evidence type, and the authors state plainly that the assembled landscape is a qualified literature survey, not a normalized ranking1.

How it works

The core mapping is between linear algebra and wave scattering. A structure's scattering matrix relates incoming to outgoing waves; design the structure and you have instantiated a linear operator that acts on an input wavefront at the speed of propagation, with no clock and, in a passive device, no standby dissipation. The review's second pillar is feedback: to solve an equation such as Ax = b, wrap the operator in a loop so the system relaxes to a fixed point; when the loop converges, the field state equals the inverse operator applied to the input. Convergence is governed by spectral radius: the Neumann series for the resolvent converges if the operator norm is below one, and more generally if the spectral radius rho(K) is below one, with the error contracting geometrically as K to the power n. The authors also flag the nonnormal loophole: matrices with spectral radius below one can still transiently amplify error, so asymptotic convergence guarantees do not rule out large intermediate excursions1.

The third pillar is the toolchain. Inverse design treats Maxwell's equations as the constraint and the geometry as the optimization variable, using adjoint methods and fabrication-aware regularization; scientific machine learning reduces the cost of repeated redesign and enables in-situ training where the hardware's own response, drift included, is what the optimizer sees. This matters because it is the practical answer to the standard objection that analog hardware is unprogrammable: you do not program the device after fabrication, you train around its imperfections, and increasingly you train inside them1.

The experimental survey is deliberately heterogeneous. Spatial differentiators and equation solvers appear at microwave through optical frequencies; integrated platforms include Mach-Zehnder meshes as canonical programmable linear processors; reservoir-style and recurrent architectures add memory; and the performance table records, side by side, a digital A100 at 312 TOPS and 0.78 TOPS/W on dense-FP16 peak over board TDP, a photonic PACE accelerator at 8.19 TOPS and 2.38 TOPS/W including laser power, a free-space S2NN at 417.96 TOPS and 12.59 TOPS/W against a 33.2 W component budget, a thin-film lithium niobate random-vector multiplier at 0.11 TOPS/W, a next-generation reservoir engine at 211 TOPS and 41 TOPS/W, and the free-space ACCEL system at 4550 TOPS and 74800 TOPS/W, with the explicit footnote that ACCEL's count assigns more than 99 percent of operations to optics on a three-class ImageNet task and is not a general matrix-multiplication benchmark1.

Where a skeptic should push

The review's own caveats do most of the skeptical work, which is a sign of quality. The headline efficiency spread, from 0.78 to 74800 TOPS/W, spans four orders of magnitude precisely because the denominators are incomparable: vendor peak versus wall-plug measurement, equivalent optical operations versus floating-point operations, component budgets versus full-apparatus power. The authors' stated position is that ratios across these reporting bases do not demonstrate application-level advantage. A reader who quotes any single photonic efficiency number from this paper without its boundary annotation is misreading it.

The deeper pressure point is the same one that has followed analog computing for fifty years. The review lists optical loss, analog noise, limited precision, hardware variability, and the energy and latency overhead of converting data between electronic and optical domains as standing obstacles, and adds the difficulty of nonlinear operations and dynamic reconfigurability. Most demonstrated metadevices are linear or weakly nonlinear, fixed after fabrication or slowly tunable, and evaluated on tasks chosen to suit them. The skeptics' question is unchanged: where, on a full application datapath including conversion overhead, does a wave computer beat a GPU that simply absorbs more watts?

Finally, weight the evidence type. This is a review; the experimental numbers are quoted from primary sources with their original assumptions, several of them calculated component budgets rather than measured wall-plug efficiencies, and the review marks them as such. That is the right epistemic posture, but it means the survey establishes that a matheatical and experimental program exists, not that any metadevice wins a head-to-head workload today.

Wave computing closes in on the organoid niche

The uncomfortable news for organoid intelligence is that the photonics community now occupies much of the rhetorical ground the organoid field considers its own. Computation delegated to exotic physics; linear algebra performed by propagation rather than by instruction fetch; training performed in situ around hardware imperfections; energy budgets quoted against biological baselines. The slogan that you should let the substrate do the math is being executed in guided lithium niobate and free-space diffractive stacks, and it does not require a perfusion rig, a contamination strategy, or an ethics opinion1.

The honest response is to note where the two substrate programs diverge, because the divergence is exactly where the organoid case lives. Wave metadevices are superb at fixed or slowly reconfigured linear operators and notoriously bad at cheap, local, state-dependent nonlinearity; tissue is the inverse. A neuron's conductances are memductive elements with memory, gating and adaptation as native physics, available at every synapse, whereas photonic nonlinearities demand special materials and intensities that the review notes must stay below damage thresholds. If your workload is a big linear transform, waves will likely beat tissue on bandwidth, reproducibility and shelf life. If your workload needs plasticity, chemical state, and self-repair, the comparison has not even started, because photonics has no native mechanism to offer.

The second, and arguably more valuable, implication is methodological. This review's reporting discipline, explicit power boundaries, identified reporting bases, component budgets separated from wall-plug measurements, equivalent-operation conventions flagged as such, is precisely the accounting the organoid-computing literature avoids. The field still quotes twenty-watt brain-versus-datacenter comparisons without stating the boundary. An organoid energy claim published to the standard of this review's Table 2 would have to state whether it counts perfusion pumps, incubation, staff time, and the energy cost of weeks of differentiation amortized over how many inferences. Adopting that discipline would shrink the headline numbers and, in exchange, make them defensible. The metadevice community is solving the same calibration-drift and hardware-variability problems that bedevil every electrode-tissue interface, and its inverse-design toolchain, treat the physical system as it measures, not as it was designed, is directly transferable to characterizing heterogeneous organoids1.

The threat is real but narrower than the hype on either side: photonics commoditizes the linear-algebra layer that some organoid proposals implicitly rely on, and it sets a documentation bar that organoid papers will increasingly be judged against. The opportunity is that the bar comes with a template.

The bottom line

Established as a synthesis: the mathematical mapping from linear operators to wave scattering, from equation inversion to feedback loops with spectral-radius convergence conditions, and from design objectives to fabricable structures via inverse design, is coherent and well documented, with experimental demonstrations across four frequency regimes. Not established: any normalized, application-level energy advantage of wave computing over digital electronics, which the review itself declines to claim. For organoid intelligence the actionable takeaway is twofold: stop competing with physics-based analog computing on its home turf of linear transforms, and adopt its reporting discipline before its standards get imposed on you. What would change this picture is a head-to-head benchmark on an identical workload and identical accounting boundaries, across a photonic processor, a GPU, and an organoid system. Nobody has run it.

Frequently asked questions

What is a metadevice?

The review's collective term for metamaterials, metasurfaces and integrated metastructures whose geometry and composition are engineered so that wave propagation through them performs a mathematical operation, such as a linear transform, a differentiation, or a step in solving an equation.

How can waves solve equations?

By feedback. The desired operator is embedded in a loop whose fixed point corresponds to the equation's solution; if the loop operator's spectral radius is below one, the iteration converges and the steady-state field equals the inverse operator applied to the input. Convergence can still suffer transient error growth for nonnormal operators.

Are photonic computers more energy-efficient than GPUs?

The review declines to say. Quoted efficiencies range from 0.78 TOPS/W for an A100 on dense-peak-over-TDP accounting to 74800 TOPS/W for a free-space system whose operation count assigns over 99 percent of work to optics on a narrow task. The numbers use different reporting bases and power boundaries and cannot be ratioed into a ranking.

Why does this review matter to organoid computing?

Photonics is pursuing the same core idea, delegating computation to physical substrate dynamics, without living tissue, and its documentation standards are stricter. It competes directly on linear-algebra workloads and indirectly by setting the accounting discipline that all unconventional-computing energy claims will be judged against.

Where does living tissue still have an edge?

In native nonlinear, plastic, chemically coupled dynamics. State-dependent conductances with memory exist at every synapse in tissue, while photonic nonlinearities require special materials and constrained intensities. Workloads that need adaptation, self-repair and chemical context are the ground on which the two substrates have not genuinely been compared.

References

  1. Hadia, X., Tsilipakos, O., Tzarouchis, D. C., Pestourie, R. Metadevices that compute: mathematical foundations and experimental implementations for neuromorphic nanophotonics. arXiv:2610.40903 [physics.optics], 2026. https://arxiv.org/abs/2610.40903. Accessed 2026-11-17.