Structure Enforced by Wiring: Reversible and Dissipative Operator Splitting on Noisy Analog Crossbars

Learned physics simulators predict a system by applying a learned update step over and over. When the hardware computing each step is imprecise, the errors pile up. Analog in-memory hardware is imprecise by design. It stores network weights as electrical conductances in a grid of memory cells, called a crossbar, and multiplies a vector by a matrix in one physical step by summing currents. Its programmed weights are off by about a percent and drift over time. For a classifier that makes one prediction, this costs a little accuracy. For a simulator rolled forward thousands of times, each small error can quietly add energy to a fluid, and the rollout blows up.

We propose to split each physics update into two kinds of sub-operator and build each one so that its physical character survives hardware error. The first kind is a time-reversible, energy-conserving part, such as advection (transport by a flow) or dispersion (waves spreading at different speeds). The second is a time-irreversible, energy-removing part, such as diffusion or damping. The reversible part uses a matrix of the form $A - A^\top$, which is always skew-symmetric (it equals minus its own transpose) and so cannot change energy. The dissipative part uses $BB^\top$, which is always positive semi-definite and so can only remove energy. Both are computed from a single programmed crossbar that is read once forward and once transposed. Because the same imprecise conductances appear in both reads, the skew-symmetry and positivity hold exactly whatever the write error. Hardware noise can then make the simulation less accurate, but it can no longer create energy.

Core Claim

In a compositional simulator, conservation and dissipation guarantees can be moved from the learned weight values, which analog hardware corrupts, into the way the weights are read, which the hardware does not corrupt. Concretely, a split step of the form

$$x_{t+h} = \Phi^{\mathrm{rev}}_{h/2}\circ \Phi^{\mathrm{diss}}_{h}\circ \Phi^{\mathrm{rev}}_{h/2}(x_t)$$

has two parts. In it, $x_t$ is the discretized field at time $t$. $\Phi^{\mathrm{rev}}$ integrates $\dot x = (A - A^\top)x$ with a time-symmetric scheme such as leapfrog. $\Phi^{\mathrm{diss}}$ integrates $\dot x = -BB^\top x + f(x)$, where $f$ is a small local nonlinear reaction term. This step keeps its energy behavior under arbitrary errors in the stored matrices $A$ and $B$. For the continuous linear part, the energy $E = \tfrac12\|x\|^2$ obeys

$$\frac{dE}{dt} = -\|B^\top x\|^2 \le 0 .$$

The skew term contributes exactly zero, since $x^\top (A-A^\top) x = 0$ for any matrix $A$, perturbed or not.

The leapfrog discretization does not conserve $\|x\|^2$ exactly. As a symplectic, time-symmetric scheme on a linear skew system, it conserves a nearby quadratic quantity and shows no secular drift. We will measure how close that quantity is rather than assume it. The prediction that follows is sharp. When conductance error is swept upward, the unconstrained split simulator should show a growing fraction of steps that inject energy. The structured one should show essentially none, and should degrade only in trajectory accuracy.

Relation to Prior Work

Two compositional simulators supply the splitting machinery. DISCO (Test-time Generalization for Physics through Neural Operator Splitting) trains neural operators on isolated single-physics equations. At test time it searches for the right combination and chains them with Lie or Strang splitting, reaching an NRMSE of 0.066 on unseen 2D Navier-Stokes. Origo (Origo: Interpretable Multi-physics PDE Foundation Model through Neural Operator Splitting) splits along linear versus nonlinear lines, with a spectral linear operator and a bank of local nonlinear operators inside a symmetric Strang skeleton. Both split by equation or by linearity, not by the direction of time. Both run on exact digital arithmetic, where an advection operator that slightly adds energy goes unnoticed. We keep their splitting skeleton. We change what the pieces are forced to be.

Two lines motivate the reversible/dissipative split itself. Stable Port-Hamiltonian Neural Networks writes dynamics as a skew-symmetric conservative term plus a positive semi-definite dissipative term. This guarantees stability for small lumped systems. However, its convex energy function restricts it to a single equilibrium, which rules out Allen-Cahn phase separation and turbulence. We borrow only the $J$/$R$ algebra, applied per sub-operator on spatial fields, and leave the nonlinear local reaction unconstrained so that multiple equilibria remain possible.

Time-Reversal Symmetric ODE Network shows that a soft time-reversal loss filters noise, cutting Duffing-oscillator MSE from 50.81 to 13.78. It also has three costs: it needs a hand-specified reversing operator, it doubles training time, and it enforces symmetry only on average. Our construction needs no reversing operator, adds no training passes, and holds per step. Hamiltonian Neural Networks enforces conservation exactly but fails on damped systems. Splitting off the dissipative piece is how we avoid that failure.

Evidence on the hardware side constrains the design. Open-loop analog programmable electrochemical memory array reports electrochemical memory (ECRAM) cells with 1.07% open-loop write error and an asymmetry non-linearity of 0.21. It evaluates them only on classification. Transition from synaptic simulation to nonvolatile resistive switching behavior based on an Ag/Ag:ZnO/Pt memristor documents filament-based devices with an on/off ratio of about $10^2$, which implies coarse and variable conductance levels. These figures set the noise model we must survive. Work on long-horizon surrogates shows why the noise matters. Learned Simulators for Turbulence finds drift in conserved quantities despite noise-injection training, and ACE2 needs an explicit per-step corrector to conserve mass.

Research Program

We build a split simulator whose sub-operators are learned discretized differential operators on periodic 1D and 2D grids. Reversible sub-operators are parameterized as $A - A^\top$. Dissipative ones are parameterized as $BB^\top$, with $B$ constrained to be local, like a stencil. Local nonlinear terms, such as the reaction in Allen-Cahn and Gray-Scott, are small pointwise networks run digitally. They are a minor share of the compute.

Each operator is first trained in floating point on single-physics data, following DISCO. It is then mapped to a simulated crossbar. The simulator models write error, an asymmetric nonlinear update, conductance drift over time, and read noise, calibrated to the reported ECRAM and Ag:ZnO figures and swept well beyond them.

The key hardware assumption is that the same array can be read with inputs applied on columns and outputs sensed on rows, giving $A^\top v$ from the same conductances. We will check this assumption against the three-terminal ECRAM read path. If it fails for ECRAM, we will state that the method applies only to two-terminal arrays.

Navier-Stokes is the hard case, because its advection operator depends on the state. There we write the advection term in its standard energy-conserving skew-symmetric form as a sum $\sum_k u_k (A_k - A_k^\top)$ over velocity components. This costs one crossbar per component. We report whether the added array count is still worthwhile. This is where the approach is most likely to break, and we will say so if it does.

Baselines

The naive reference point is DISCO-style unconstrained splitting. It uses the same sub-operator decomposition, the same data, and the same parameter count, but the sub-operators are free matrices. It is evaluated first in float32, which gives the accuracy ceiling, and then on the same simulated crossbar. This tells us how much analog noise hurts when nothing protects against it.

The strongest existing method for accuracy on these tasks is operator splitting in the DISCO/Origo style. The best reported figures in the file are DISCO's NRMSE of 0.066 on unseen 2D Navier-Stokes and Origo's 50–60% relative-error reduction over MOE-POT on held-out KdV and Allen-Cahn. We will retrain a DISCO-style model on our splits at matched compute and tuning budget, rather than quoting these numbers across setups. Origo's spectral operator relies on Fourier transforms, which do not map naturally onto a crossbar, so it serves as a digital reference only. We found no reported result on Lacuna for any operator-splitting simulator on analog or noise-perturbed hardware. The analog comparison therefore has no prior number to beat.

The hardware-aware competitor is the unconstrained split simulator trained with injected conductance noise at the level it will face in deployment. This is the obvious thing a practitioner would do. We also run the per-step corrector idea from ACE2: after each step, rescale the state digitally so that it matches the energy the physics prescribes. A third competitor is the soft time-reversal loss of TRS-ODEN applied to the unconstrained split model, with the reversing operator taken as velocity sign-flip where one is known.

The ablation that isolates our key ingredient keeps the $A - A^\top$ and $BB^\top$ parameterization. However, it stores the full products $A - A^\top$ and $BB^\top$ as separately programmed matrices, so the structure holds at training time but is broken by independent write errors. If this ablation performs as well as the full method, then the benefit comes from the parameterization and not from structure through the read geometry, and the hardware claim fails.

Beating the baselines means two things. First, a fraction of energy-injecting steps statistically indistinguishable from zero at every noise level. Second, rollout NRMSE at step $10^4$ no worse than the noise-aware baseline, with float32 accuracy within about 10% of the unconstrained model.

Why not just…?

Why not just train the unconstrained model with hardware noise injected? This is the most direct threat. Noise-aware training teaches the model to tolerate noise on average. It cannot guarantee the sign of the energy change on any given step, and drift changes the conductances after training anyway. Learned Simulators for Turbulence reports that its noise-injection trick stopped blow-up but not drift in conserved quantities. The drift experiment below settles the question directly. If noise-aware training matches our energy behavior after simulated conductance drift, the structural approach adds nothing.

Why not just apply a conservation corrector after every step, as ACE2 does? A global rescaling fixes one number, the total energy. It does not stop energy moving between spatial scales or regions, so spectral error can still grow even when the total is correct. It also needs digital computation of a global norm every step, which cuts into the analog speed advantage. It is a real competitor on global quantities. We therefore report the Fourier spectral error (fRMSE) and the conservation error (cRMSE) separately, and count the corrector's digital operations.

Why not just use closed-loop write-and-verify programming, or stay digital? Write-and-verify shrinks the programming error. It does not eliminate it, and it does nothing about later drift. The open-loop ECRAM result exists precisely because verification is slow. Staying digital is a legitimate choice where energy budgets allow. Our claim is limited to the regime where analog is used at all. We make no speed or energy claims that we do not measure in simulation.

Why not just use a soft time-reversal loss? The loss acts during training and on average. It cannot constrain errors that the hardware introduces at deployment, and it needs a reversing operator that is not defined for dispersive or reaction terms.

Proposed Experiments

We evaluate on 1D advection and reaction-diffusion from PDEBench, the 1D combined advection-diffusion-dispersion setting and 2D Gray-Scott from DISCO, 1D KdV and 2D Allen-Cahn as held-out equations following Origo, and 2D incompressible Navier-Stokes as the stress test. Metrics are nRMSE, PDEBench's cRMSE and fRMSE, the energy error relative to ground truth, and the fraction of steps where energy rises in regimes where the physics forbids it.

The first decisive check sweeps write error from 0 to 5% and measures that energy-injection fraction for every model. The theory predicts zero for our method and for no other model. The second runs $10^4$-step rollouts while simulated conductance drift accumulates, without reprogramming. It tests whether our model degrades gracefully while the noise-aware baseline eventually diverges. The third tests zero-shot composition on KdV and Allen-Cahn: operators are trained only on isolated pieces, and the check is whether the reversible/dissipative typing costs accuracy relative to DISCO-style free operators.

The proposal fails in any of three cases: the separately-stored ablation matches the full method, noise-aware training matches it after drift, or float32 accuracy on held-out equations drops by more than about 10%. A small demonstration on a physical $10\times10$ ECRAM array for a 1D advection-diffusion problem would test the transpose-read assumption directly, if such an array is available.