Analog "dynamical system" computers solve equations by letting a physical circuit evolve in continuous time instead of stepping a digital clock. They promise large gains in speed and energy, but the published evidence has two holes. First, nearly all of it comes from circuit simulation rather than hardware. Second, the designs assume every node talks to every other node, so the wiring grows with the square of the problem size. The standard fix for that wiring cost, splitting busy nodes into strongly tied "copies", was validated only for problems that settle to a static answer. Nobody has asked what it does to a solver that must track a trajectory as it moves. This note proposes answering that question on a system whose ground truth is measured, not simulated: the fluorescence a leaf emits in the first second after a bright flash. The curve is cheap to record, it is shaped by a sparse chain of electron carriers, and it spans five decades of time, from microseconds to hundreds of milliseconds.
The core claim is simple. For a continuous-time analog solver, the right sparse wiring is the system's own interaction graph. Copy-node sparsification, which works for static optimization, adds a trajectory error set by how long the copies take to agree compared with the fastest timescale in the system. On a stiff, multi-timescale system this error becomes a visible phase lag or amplitude distortion. We will test the claim in circuit simulation and on a small physical board, and score both against real transients recorded from stressed plants. As a byproduct, this gives the first measurement of how far simulation and hardware diverge for this class of solver.
Why this intersection, and what is already done
The analog solver we build on is DS-TS from Solving Time-Dependent Differential Equations with Physical Dynamical Systems. It solves time-dependent equations with coupled analog integrators and conductances quantized to 8 bits, and it reports mean absolute error one to three orders of magnitude below learned solvers such as physics-informed networks, Fourier neural operators and UNets. It also reports roughly 1.12 W against 250 W for an A100. All of these results come from simulation. The same authors note that the all-to-all coupling leads to a "resistor explosion", meaning the number of resistors grows until the chip cannot be built, and they leave sparse topologies as future work (Liu et al.). DS-TS supplies our method. We keep its architecture and change only its wiring.
The sparsification technique comes from Scalable connectivity for Ising machines: Dense to sparse. An Ising machine is hardware that searches for low-energy states of a network of interacting binary units. That paper clones high-degree nodes into copies joined by strong couplings of weight $W_0$. In layout simulation this cuts area scaling from $O(N^{1.03})$ to $O(N^{0.34})$, where $N$ is the number of nodes. The cost is a convergence exponent that rises from about 0 to about 4, and a $W_0$ that must be tuned by hand for each graph. The same paper notes that problems which are sparse from the start avoid these overheads altogether. We use it in two roles. It motivates the question, and its method is the main sparsification competitor.
Two other hardware efforts show the same simulation-heavy pattern. BRIM, a resistively coupled analog Ising machine, and the ferroelectric-transistor latch design of Mallick et al. both test large systems only in simulation. Their physical prototypes have 6 and 10 nodes. Both also target static MaxCut solutions, not moving trajectories.
On the plant side, existing simulators sit at two extremes. Data Augmentation From RGB to Chlorophyll Fluorescence Imaging generates fluorescence images as static textures plus Gaussian noise and needs real images to close the gap. Efficiency of energy funneling in the photosystem II supercomplex runs exact quantum dynamics for 93 pigments on GPUs, never compares against leaf transients, and cannot scale to tissue. Its value here is structural. It shows that the pigment and carrier network is local and chain-like: outer antenna, then inner antenna, then reaction center.
Finally, Chlorophyll Fluorescence Imaging for Environmental Stress Diagnosis in Crops shows that simple two-wavelength fluorescence ratios diagnose combined heat and moisture stress poorly, with $R^2 < 0.3$, and that a supervised linear classifier on spectra does better. It gives us a stress design to copy and a realistic diagnostic baseline.
No paper in the file connects analog connectivity scaling to fluorescence kinetics. The obvious version of that connection, "accelerate photosynthesis models with analog hardware", is not worth doing. A reduced model of the electron transport chain has tens of states, and a standard stiff solver on a laptop handles it trivially. We therefore use the plant system as a hard, real-data test bed for a hardware question, and we claim only a secondary application benefit.
Core Claim and Mechanism
The recorded curve is called the OJIP transient, after its four landmark steps. It rises from an initial level at tens of microseconds, through inflections near 2 ms and 30 ms, to a peak near 200–500 ms. These steps reflect the successive filling of the chain's electron acceptors. A learned emulator of this curve can be wired in three ways.
The first is dense, with all-to-all learned coupling, as in DS-TS. The second is copy-node, where a dense emulator is split into copies with the degree capped at $k$. The third is native sparse, where coupling is allowed only along the carrier chain within each tissue pixel, plus nearest-neighbor coupling between pixels to capture lateral heterogeneity. The third option is what makes the quadratic wiring problem real. A 64×64 fluorescence-imaging patch with about 20 states per pixel is roughly $8\times10^4$ nodes, which a dense design cannot wire.
The mechanism behind our prediction is a separation of timescales. Copies tied by a coupling conductance proportional to $W_0$ on node capacitance $C$ pull each other into agreement with time constant $\tau_c \approx C/W_0$. A standard singular-perturbation argument, which treats the copies' disagreement as a fast mode relaxing on $\tau_c$, gives
$$\epsilon_{\text{copy}}(t) \sim \frac{\tau_c}{\tau_{\min}(t)},$$
where $\epsilon_{\text{copy}}$ is the relative disagreement between copies and $\tau_{\min}(t)$ is the fastest active physical timescale at time $t$. For OJIP, $\tau_{\min}$ is microseconds in the first rise and hundreds of milliseconds by the peak. We therefore predict that copy-node error concentrates in the fast early phase and appears as a lag in the time of the 2 ms step.
Raising $W_0$ enough to suppress this error adds stiffness and increases sensitivity to noise and to 8-bit quantization. That is the continuous-time analogue of the freezing trade-off seen in static optimization. Native sparse wiring avoids the trade-off because no artificial agreement constraint exists.
Research Program
We will build three analog emulators that share one architecture. Each keeps DS-TS's excitatory-inhibitory coupling, its state-dependent nonlinearity and its 8-bit quantization-aware training, and differs only in wiring: dense, copy-node, or native sparse. Each is trained to reproduce measured OJIP transients, conditioned on plant stress state through a small set of per-pixel parameters.
We will evaluate them at three levels. At the first level, a numerical reference integrates the same equations with a stiff solver. At the second, SPICE circuit simulation adds thermal noise, transistor mismatch and quantization. At the third, a physical printed-circuit board implements one pixel's native-sparse chain of about 20 states with discrete components. This is feasible precisely because the native graph is small and local, and it is the step no prior analog trajectory solver has taken.
The difference between the second and third levels measures the simulation-to-hardware gap directly. The difference between every level and the measured leaf curves measures the model-to-reality gap.
Ground-truth data will come from a greenhouse campaign that mirrors the five-group design of the garlic study: a control, two heat levels, and two combined heat-and-moisture levels. We will record whole-leaf OJIP curves with a handheld fluorometer and spatially resolved transients with a fluorescence imager, on a leafy crop and on Arabidopsis. Using Arabidopsis ties the imaging arm to Real-Fluo. No public OJIP time-series dataset appears in the file, so collecting this data is part of the contribution, and the dataset will be released.
As a control, we will also run all three wirings on DS-TS's five-equation benchmark suite: advection-diffusion, wave, reaction-diffusion, Fokker-Planck and a higher-order hyperbolic equation. This separates effects of wiring from effects of the plant data.
Baselines
The naive reference is the scalar maximum quantum yield of photosystem II, $F_v/F_m$, which is computed from just the starting and peak fluorescence levels. For stress diagnosis we pair it with the two-wavelength ratio indices. The garlic study reports $R^2 < 0.3$ for those ratios under combined stress (Chlorophyll Fluorescence Imaging for Environmental Stress Diagnosis in Crops), although on spectra rather than transients. Any trajectory-based method must beat these ratios on our own data before it is interesting.
The strongest diagnostic competitor is standard normal variate preprocessing followed by partial least squares discriminant analysis (PLS-DA), which was the best method in that study. We will apply it to the raw sampled OJIP curves, using the same training and test plants and the same cross-validation folds as our emulators. Beating it means higher held-out stress classification accuracy, or equal accuracy with physically interpretable per-state parameters.
The strongest competitor on the hardware question is the dense DS-TS emulator (Liu et al.). It reports MAE one to three orders of magnitude below learned solvers on its five-equation suite, at about 1.12 W, but no reported result was found on Lacuna for any photosynthetic or fluorescence task. We will train it with the same data, the same number of optimization steps, the same hyperparameter search budget and the same 8-bit quantization. Native sparse wiring "beats" it if trajectory error is within noise of the dense emulator while using orders of magnitude fewer coupling elements. At the 64×64 patch scale, where dense wiring cannot be laid out, beating it simply means working at all.
The copy-node baseline follows Sajeeb et al., which reports area scaling of $O(N^{0.34})$ and a convergence exponent of about 4 on static MaxCut. We give it an explicit sweep over $W_0$ and over the degree cap $k$, with a tuning budget equal to that of the native model. We report its best setting rather than a strawman.
A digital baseline is a time-multiplexed field-programmable gate array (FPGA) emulator in the style of the FPGA cyber coherent Ising machine, running the native graph. No reported result was found on Lacuna for it on trajectory tasks. A stiff numerical solver on a CPU serves as the accuracy ceiling for the model equations themselves.
The ablation that isolates the key ingredient keeps the native-sparse coupling count fixed but rewires the connections randomly, using bounded-degree random graphs of the kind studied by De Santis et al.. If random sparse wiring matches native wiring, then the claim reduces to "any sparsity works", and the biological graph is not what matters.
Why not just…?
Why not just integrate the kinetic model with a standard stiff solver on a CPU? For plant science alone, that is the right answer, and we concede it. A 20-state model does not need an accelerator, and the CPU solver is our accuracy ceiling, not a competitor we expect to beat. The proposal's claim concerns analog trajectory solvers. The plant system is chosen because its native graph is sparse, its time range is extreme, and real ground truth can be measured cheaply. Throughput matters only in the imaging arm, where every pixel needs its own parameter fit. There we will report wall-clock time and energy per fitted image against a batched GPU solver, and we will state plainly if the GPU wins.
Why not just test sparse wiring on the existing PDE benchmarks, since discretized diffusion is already locally coupled? This is a genuine threat, and it is why the five-equation suite is our control. If copy-node and native wiring behave identically there and on OJIP, the plant data adds only realism, not new insight. Our timescale argument predicts a larger copy-node penalty on OJIP than on the PDEs, because OJIP's fastest timescale is five decades shorter than its slowest. The benchmarks also have synthetic ground truth, so they cannot measure the gap between model and reality.
Why not just time-multiplex a digital FPGA and skip analog wiring entirely? Time-multiplexing removes the wiring problem but reintroduces discrete time steps. DS-TS argues that those steps are the source of the trade-off between latency and fidelity. Our FPGA baseline tests that argument directly on a stiff real system. If the FPGA matches the analog board's error at comparable energy, the case for analog trajectory solvers here weakens, and we will report that result as such.
Why not just classify stress with PLS-DA on the curves and skip dynamics? PLS-DA may well classify as accurately as our emulators, and it is the diagnostic baseline. What it cannot provide is per-carrier parameters that say where the electron transport chain is failing. We will count the diagnostic arm as a success only if emulator-derived parameters match PLS-DA's accuracy and also separate heat stress from combined heat-and-moisture stress in a way consistent with the physiology.
Proposed Experiments
The decisive experiment sweeps $W_0$ and the degree cap $k$ for the copy-node emulator and measures error in the timing and amplitude of each step against the native-sparse emulator. The primary metric is error on a log-spaced time grid, evaluated against measured curves. The prediction is that copy-node error scales with $\tau_c/\tau_{\min}$, concentrates before 2 ms, and cannot be driven below native-sparse error without breaking under 8-bit noise. If copy-node error is flat across $W_0$, the timescale argument is wrong.
The second experiment compares the physical board with its SPICE twin, using identical trained parameters on the same leaf curves. It reports the added error, the drift over temperature, and the fraction of the simulated advantage that survives in hardware. No prior dynamical-solver paper in the file reports this number.
The third experiment scales native-sparse wiring from a single pixel to a 64×64 patch in simulation. It reports coupling-element count and estimated area against dense wiring, to check whether native sparsity reaches the favorable area exponent Sajeeb et al. found for copy nodes without paying their convergence penalty.
The fourth experiment compares held-out stress classification from emulator parameters, PLS-DA, $F_v/F_m$ and ratio indices across the five-group design.
References
- Solving Time-Dependent Differential Equations with Physical Dynamical Systems
- Scalable connectivity for Ising machines: Dense to sparse
- BRIM: Bistable Resistively-Coupled Ising Machine
- CMOS-compatible ising machines built using bistable latches coupled through ferroelectric transistor arrays
- Data Augmentation From RGB to Chlorophyll Fluorescence Imaging Application to Leaf Segmentation of Arabidopsis thaliana From Top View Images
- Efficiency of energy funneling in the photosystem II supercomplex of higher plants
- Chlorophyll Fluorescence Imaging for Environmental Stress Diagnosis in Crops
- Highly Versatile FPGA-Implemented Cyber Coherent Ising Machine
- Low degree connectivity of ad-hoc networks via percolation