Scheduling Polarity Swaps in Solar-Driven Symmetric Electrolyzers

A water-splitting cell produces hydrogen at one electrode (the cathode) and oxygen at the other (the anode). Some newer catalyst materials, called bifunctional catalysts, can run either reaction, so the same material can be placed on both electrodes. That design is usually sold on cost and simpler manufacturing. We think it offers something more useful. In a cell with identical electrodes, the operator can reverse the current, and the two electrodes then trade jobs. Polarity becomes a control input. The electrode that has been oxidizing and wearing down as the anode can spend time being reduced as the cathode, and the other electrode takes over the harsh anode duty.

Whether reversal heals or harms the catalyst depends on the catalyst's hidden surface state and on when the swap happens. We propose to decide swaps and current levels with receding horizon control. This is a controller that, every few minutes, predicts the next stretch of time from a model and a solar forecast, picks the best plan, applies only the first step, and then re-plans. Its model tracks an estimated "how oxidized and how worn" state for each electrode, updated online from voltage signals the cell already produces. Solar power has natural dips and ramps, and the controller should place swaps into those cheap moments while keeping each electrode's surface inside a safe window.

Core Claim

We claim that a symmetric alkaline cell with the same bifunctional catalyst on both electrodes can have its polarity and current scheduled by a forecast-aware receding-horizon controller, and that this controller will lose activity measurably more slowly under real solar profiles than fixed-rule operation, at equal or higher hydrogen output. The gain has to come from two ingredients working together. The first is an online estimate of each electrode's surface state. The second is a look-ahead that times swaps and protective holds against the forecast. We design the experiments so either ingredient can be removed and its contribution measured. If the simple rules described below perform as well, the claim is false, and that result is worth publishing too.

Why This Gap Is Real

The materials papers motivate the problem, and they also show where the problem sits. CoP/Fe‐Co9S8 for Highly Efficient Overall Water Splitting with Surface Reconstruction and Self‐Termination reports that its high activity comes from surface reconstruction during operation into high-valence Co(IV)–O sites. Left uncontrolled, the same reconstruction causes collapse and leaching, and the authors stop it chemically by adding iron. A Dual-Function Fe-Doped Co3O4 Nanosheet Array for Efficient OER and HER in an Alkaline Medium tunes activity by doping and reports 200 hours of stability. Both papers test under constant current holds. The PEM (proton exchange membrane) bifunctional system in Core–Shell IrPt Nanoalloy on La/Ni–Co3O4 still loses 5 µV per hour over 646 hours, also under steady operation. None of these works tests the intermittent, reversing regime that solar operation creates. None treats reconstruction as something operation could steer rather than composition alone. The trifunctional Ni–P/Fe–P system connects solar cells, a Zn–air battery and an electrolyzer with no active control at all.

On the control side, NMPC for Mode-Switching Operation of Reversible Solid Oxide Cell Systems is the closest work. It uses nonlinear predictive control to switch a high-temperature ceramic cell between electrolysis and fuel-cell modes, and it penalizes a thermal-stress quantity. Its electrodes are fixed porous media with constant kinetics, and it assumes the model matches the plant perfectly. Economic NMPC for a Reversible Solid Oxide Cell adds electricity-price arbitrage and keeps the same fixed-catalyst assumption. Degradation-aware data-enabled predictive control of energy hubs shows that putting a wear penalty inside a look-ahead controller halves battery degradation compared with rules. Its wear model is an empirical cycle count for batteries. We borrow three things from this literature: the pattern of penalizing a stress state, the wear penalty itself, and the online model updating of EKF-SINDy-MPC. What is new here is the controlled object: a catalyst's hidden surface state, with polarity reversal as an actuator. No system in these papers has that actuator, because none of them has identical electrodes.

Research Program

The first phase builds a small grey-box model of each electrode. Grey-box means physically structured, with parameters fitted to data. For electrode $i$ it carries two hidden quantities. The first, $r_i$, is the fraction of the surface in the reconstructed active phase; it grows under anodic polarization and relaxes under cathodic polarization or at rest. The second, $d_i$, is irreversible loss from leaching or delamination, and it grows when $r_i$ leaves a safe band or when the potential changes quickly. Measured cell voltage is written as a function of current, $r_i$, $d_i$, and a bubble/resistance term. We identify the model's terms with sparse regression, using the SINDy (sparse identification of nonlinear dynamics) approach from EKF-SINDy-MPC. The training data come from designed excitation runs that mix current steps, rest periods and reversals. An extended Kalman filter, a standard recursive estimator for nonlinear models, then updates $r_i$, $d_i$ and slowly drifting parameters online. Its inputs are the voltage trace and the short relaxation transient seen at each swap. We add a brief impedance measurement every few hours to check the estimated charge-transfer resistance against a direct reading.

The second phase is the controller. Every sampling period $\Delta t$, it takes the current state estimate and a solar forecast over horizon $N$. It then chooses current $j_k$, polarity $p_k \in \{+1,-1\}$ and battery power $b_k$ to solve

$$\min \sum_{k=0}^{N-1} \Big[ -\alpha\, H(j_k) + \beta \sum_i \Delta d_{i,k} + \gamma\, \mathbb{1}[p_k \ne p_{k-1}] \Big] \quad \text{s.t. } r_{i,k} \in [r_{\min}, r_{\max}],\; |j_k - j_{k-1}| \le \delta,\; \text{power balance with solar and battery}.$$

Here $H$ is the hydrogen produced and $\Delta d_{i,k}$ is the predicted irreversible loss. The indicator term is a fixed cost per swap, covering the gas-crossover purge and the lost production. The weights $\alpha,\beta,\gamma$ trade these terms against each other. The binary polarity choice makes the problem a small mixed-integer program. Over a one-hour horizon at one-minute steps it has a few dozen binaries, which we expect to solve well within the sampling window. The rSOC controller solved a far larger 47,000-equation problem in 2.4 CPU seconds on a 75-second sample time (Li et al.). The forecast comes from a model of the Solar Irradiance Anticipative Transformer type. We handle forecast error with a few scenarios, following Stochastic MPC for energy hubs. That paper found that moving from 1 to 100 scenarios cut constraint violations from about 30% of the time to nearly zero.

The third phase is bench validation on real cells driven by replayed solar traces. The main test catalysts are Fe–Co3O4 on nickel foam and CoP/Fe–Co9S8, each built as a symmetric alkaline cell. We chose them because their static performance and their reconstruction behavior are both reported.

Baselines

The naive reference is the cell following the available solar power directly: no swaps, no battery shaping, and shutdown whenever power falls below the minimum operating current. This is how the passive system in Ni–P/Fe–P behaves, and it shows how much damage intermittency causes before any control is applied.

The strongest non-predictive competitor is a tuned rule controller. It combines three rules: a clamp on how fast the current can ramp, a protective minimum current held from the battery during dips so the anode never falls to open-circuit potential, and a voltage-triggered polarity swap whenever measured cell voltage at the reference current rises by a set margin. This baseline uses feedback, it is easy for engineers to adopt, and it covers most of what the controller could exploit. A second rule baseline swaps polarity on a fixed period, with the period tuned. For fairness, every controller uses the same catalyst batch, the same battery capacity and the same replayed irradiance traces. Each rule baseline gets the same number of bench tuning trials on a held-out training trace that the predictive controller gets for its weights. All controllers run on at least three independent cells.

On the control side, the most relevant reported number is that predictive control cut annual battery capacity loss to 0.3%, against 0.8% for a rule-based controller (Degradation-aware DeePC). That result is for batteries, not catalysts. On the catalyst side, we found no reported result on Lacuna for the degradation of bifunctional catalysts under dynamic or reversing profiles. The only reference points are static: 200 hours for Fe–Co3O4/NF (Su et al.); 1.68 V at 1 A cm⁻² sustained for more than 150 hours for CoP/Fe–Co9S8 (Cheng et al.); and 5 µV h⁻¹ over 646 hours for the IrPt PEM cell (Liu et al.). We will count the proposal as beating the tuned rules if it achieves a clearly lower voltage-drift rate at the reference current, with confidence intervals across cells that do not overlap, at equal or higher hydrogen output per unit of solar energy. Our target, which we are not claiming in advance, is a 30% reduction.

Three ablations isolate the key ingredients. The first keeps the predictive controller but removes swapping, so polarity is fixed; this tests whether the bifunctional actuator matters. The second keeps the full controller but replaces the forecast with persistence (assume current conditions continue); this tests the look-ahead. The third freezes the state estimator at its initial parameters; this tests online estimation.

Why not just…?

Why not just clamp current ramps and hold a protective current during dips? This is the most serious threat. If damage depends only on the instantaneous potential staying above a threshold, the clamp-and-hold rule should match the controller. The rule has two costs: it spends battery energy on every dip, however short, and it cannot see how much accumulated reconstruction an electrode already carries. We run it as the main baseline with full tuning. If the no-swap ablation of our controller performs no better than this rule, we conclude that look-ahead adds nothing for fixed-polarity operation, and the case then rests on polarity scheduling alone.

Why not just reverse polarity on a fixed timer, or whenever voltage rises? Periodic reversal needs no model, and the voltage-triggered version already uses feedback. It cannot time a swap into a forecast dip, where the swap costs almost no production. It also cannot detect that a swap would push the new anode, already heavily reconstructed, past its safe band. Both versions are baselines. If either matches the full controller, forecasting and state estimation are unnecessary. That would be a useful practical finding, and the paper would report it that way.

Why not just make the material self-terminating? Cheng et al. show that chemistry can limit reconstruction. Their evidence comes from steady holds, and the same file reports residual drift even in heavily engineered systems (Liu et al.). We include CoP/Fe–Co9S8 exactly because it is the best-engineered case. If control adds nothing on it but helps Fe–Co3O4, then control and chemistry substitute for each other, and we will say so.

Why not a standard PI loop? A PI (proportional-integral) loop can track a power setpoint, but it has no way to express a swap decision or a constraint on a hidden state. Li et al. report overshoot and larger stress peaks under a cascade PI scheme compared with NMPC. We therefore treat PI only as the inner current loop inside every controller, not as a competitor.

Proposed Experiments

The first check is identifiability, done before any control comparison. Using excitation runs, we test whether the estimated $r_i$ tracks independent measurements of surface oxidation. These are double-layer capacitance as a proxy for active area, and ex situ spectroscopy on sacrificial cells stopped at chosen points. If the voltage signal cannot distinguish reconstruction from bubble resistance, the state-based controller has nothing to act on. In that case we would drop down to the rule baselines and report the negative result.

The second check establishes the reversal response. On each catalyst, we compare swap frequencies across a matrix of reversals at different accumulated anodic charge, measuring voltage drift and leached metal in the electrolyte. This tells us whether reversal regenerates the catalyst or damages it, and in which state window each happens. If reversal only harms, the polarity actuator is dead, and the proposal narrows to degradation-aware current scheduling.

The main comparison replays irradiance traces from the SIRTA and Chilbolton sites used by SIAT, compressed in time to cover about 300 hours of accelerated cycling per cell. On these traces we run the naive, rule, fixed-period, full-controller and three ablation arms. The reported outcomes are voltage drift at the reference current, hydrogen produced per unit of solar energy, change in charge-transfer resistance, loss of active area, and dissolved metal measured post mortem. The decisive result is the gap between the full controller and the voltage-triggered swap rule on Fe–Co3O4. If that gap is not clear, the central claim fails. An acidic extension on the IrPt PEM cell follows only if the alkaline results are positive, since reversal in acid risks dissolving the noble metal.