Teaching physics to generate, without a target trajectory

What the work claims

This is a primary methods-and-measurement paper. The generator of a Wasserstein GAN is replaced by an Analog Interaction System: a set of coupled physical elements, oscillators in the concrete implementation, evolving a fixed differential equation whose functional form is set by hardware physics rather than chosen by a designer. Training adjusts only what hardware can adjust, coupling weights and schedule parameters, and penalizes only the terminal visible state, letting the physics find its own path from noise to data.¹

The headline results: on MNIST and Fashion-MNIST, the best oscillator model, Kuramoto dynamics extended with superharmonic injection locking (KuraSHIL), reaches FID scores of 27.6 and 80.8 at full precision without noise, and 29.1 and 75.1 under 4-bit weight quantization with transient noise. Reproduced hardware-oriented baselines sit far behind: the diffusion Trajectory Machine scores 107.8 and 112.8, and the Neural Langevin Machine 230.5 and 200.8. The authors estimate 23 microjoules per generated image for their architecture, about two orders of magnitude better than the 7 to 79 millijoules of digital baselines they cite.¹

The deeper claim is diagnostic. Fixed physical interactions, sinusoidal coupling, fixed polynomial or sigmoidal nonlinearities, approximate a narrower class of dynamics than neural networks can express. The authors characterize that expressivity gap empirically on low-dimensional synthetic distributions and identify two hardware-compatible mechanisms that close it: piecewise time-varying parameters, which effectively raise dynamical complexity, and hidden physical elements, which expand the state the dynamics can use.

How it works

Standard diffusion and flow-matching generators integrate a neural-network vector field step by step; the analog hardware instead executes one continuous differential equation whose vector field is built from physical interactions, so inference energy is set by physics rather than by the number of integration steps. The mismatch is that the physics-imposed equation may be a poor approximator of the flexible vector field a generative model wants. Exact-trajectory supervision, as in flow matching, is infeasible for such hardware because the fixed dynamics cannot be forced onto an arbitrary path. The fix is distributional: train with a Wasserstein GAN with gradient penalty, where the discriminator compares the terminal state distribution against real data and the generator's internal trajectory is unconstrained.¹

In low-dimensional studies, the authors measure how well each AIS class matches target distributions using sliced Wasserstein distance, comparing training objectives and configurations with and without hidden states. Hidden-state augmentation consistently gives the lowest distances, and freeing the objective from trajectory constraints lets that extra expressivity be exploited. In the image-scale architecture, visible elements map to pixel dimensions on a grid interleaved with hidden elements, connected through programmable unidirectional coupling units, with weight precision and connectivity treated as design variables rather than afterthoughts.¹

The ablations are specific. Removing hidden states degrades FID from 27.6 and 80.8 to 43.0 and 84.3. Increasing the number of piecewise time segments for the weights from one to four reduces FID on both datasets. Four-bit quantized weights match the 64-bit baseline (27.2 and 81.9 against 27.6 and 80.8), while pushing transient noise from 0.025 to 0.1 degrades MNIST FID from 29.1 to 54.5. The power model is built from published silicon: a 28-nm coupled-oscillator Ising machine with 1,440 nodes, degree-8 connectivity, and 4-bit DACs consuming 460.3 milliwatts, giving a per-node baseline near 320 microwatts, a degree-dependent routing penalty, and a 2 to the q minus 4 scaling in bit width q.¹

Where a skeptic should push

The most load-bearing assumption is that hardware-implementable simulation stands in for hardware. No analog generative chip was fabricated or measured; everything runs as a differentiable numerical simulation under constraints intended to mimic hardware, with quantization and injected Gaussian noise. That is a legitimate and common methodology, but the 23 microjoule figure is an extrapolation chained to one published Ising machine's power draw, with layout-level routing effects explicitly excluded and acknowledged as a potential increase. Treat the energy number as a model output with error bars the paper does not provide, not as a measurement.¹

Second, the absolute generative quality is weak where it matters. FID 27.6 on MNIST is respectable for constrained analog hardware; FID 80.8 on Fashion-MNIST is not competitive with unconstrained digital models, which reach single digits. The paper's claim is relative, beating prior hardware-implementable analog generative models by 3 to 4 times on FID, and that comparison required re-implementing two baselines from their official code. A generous reading is that the field is young; a skeptical one is that the gap to digital is the honest headline and it is large.

Third, the training loop is digital and heavy. The GAN discriminator, the ODE solves during training, and the fine-tuning stages (15,000 generator iterations for each quantization and noise level) all run on conventional compute. End-to-end energy accounting, training included, would narrow the two-orders-of-magnitude claim substantially. The paper measures the cost of generating, not the cost of getting the physical system into a generating state. Finally, the expressivity-gap studies are low-dimensional and synthetic; whether the two identified mechanisms scale to genuinely hard distributions remains open.

Generative computing on fixed physical dynamics

Organoid intelligence makes the same wager this paper makes: that a dynamical system whose equations you did not design can be steered into useful computation. The paper is valuable to the field precisely because it refuses to romanticize that wager. It quantifies the expressivity gap between physics-imposed dynamics and software-defined ones, and it shows the gap does not close by optimism; it closes by mechanism, through parameters that vary in time and state you do not directly observe.¹

Both mechanisms have direct biological analogues, and this is the non-obvious transfer. Time-varying piecewise parameters are, in a dish, scheduled stimulation: an MEA or optogenetic protocol that changes the effective coupling regime of the network over the course of a task, giving one fixed tissue several dynamical personalities in sequence. Hidden physical elements are everything an organoid recording does not see: unrecorded neurons, glial state, neuromodulatory tone, metabolic variables. The paper's ablation shows those hidden dimensions are not garnish; removing them nearly doubled FID on MNIST. A living reservoir carries an enormous hidden state for free, which is a structural advantage over oscillator arrays, provided the experimenter learns to train through it rather than around it.

The training-objective lesson matters even more. Exact-trajectory supervision fails on fixed physics, and it fails catastrophically on living tissue, where no experimenter can force a culture to follow a prescribed trajectory at all. Terminal-state, distribution-matching training is the workable alternative this paper demonstrates: specify where the dynamics must end, not how it must travel. That is almost certainly the right objective class for organoid computing, and it reframes benchmark design: tasks should be specified as terminal distributions or terminal decisions, with the culture's internal path left free.

The threats are equally clear. One: the expressivity-gap measurements are a warning that rich-looking dynamics can be computationally narrow; if cortical organoid dynamics are closer to a fixed low-order interaction model than to a trainable vector field, task performance will plateau regardless of electrode count. Two: the energy bar is moving. Twenty-three microjoules per image, even discounted heavily for the extrapolation, plus no incubator, no perfusion, and no ethics review, is what non-living analog substrates will offer. Tissue must compete on what physics cannot easily provide, which this paper suggests is hidden state and adaptive coupling, not raw efficiency. Three: the paper's own accounting habit, separating generation energy from training energy, is one the OI field should adopt before critics impose it; an honest organoid energy budget includes life support, and honest analog comparisons include training cost on both sides.

The bottom line

Established: fixed physical dynamics can be trained as generative models by terminal-state Wasserstein GAN training, reaching FID 27.6 and 80.8 on MNIST and Fashion-MNIST, 3 to 4 times better than reproduced hardware-oriented analog baselines, with 4-bit quantized weights matching full precision and graceful degradation under moderate noise. Established: time-piecewise parameters and hidden physical states measurably close the expressivity gap that fixed physics imposes. Not established: end-to-end energy superiority once digital training is included, performance on harder data distributions, or behavior on actual fabricated analog hardware rather than constraint-mimicking simulation.

For organoid intelligence, the paper is a template for honest substrate engineering. What would confirm its lessons transfer: organoid experiments using scheduled stimulation as time-varying parameters and distribution-level terminal objectives, with hidden-state structure treated as a feature to exploit. What would break the analogy: measurements showing living dynamics are no more expressive than the simple interaction models this paper struggles with, or energy accounting that includes incubator support pushing tissue decisively above the analog bar. Either way, the field now has a quantified statement of the fixed-dynamics problem, and no excuse for ignoring it.