University of Waterloo Corrosion Model Finds a Battery Electrolyte Tradeoff

University of Waterloo researchers tested 20 alkaline electrolyte conditions, then used AI to locate a compromise between aluminum corrosion rate and electrochemical stability. The University of Waterloo corrosion model selected approximately 1.35 molar potassium hydroxide, or KOH, as its preferred point. Yet that result is not a recipe for a finished battery. It is a model-generated balance between two corrosion measurements under tightly controlled laboratory conditions.

That distinction creates the central tension. Aluminum-air batteries offer high theoretical energy per unit of active metal, but alkaline electrolytes also consume the aluminum when the cell is idle. Stronger alkalinity can support electrochemical activity while accelerating unwanted dissolution and hydrogen production. Selecting an electrolyte therefore means managing competing objectives, not simply minimizing one number.

The team replaced part of that trial-and-error process with an artificial neural network and two genetic optimization methods. Its study, published July 10, 2026, in Frontiers, reports prediction scores above 0.999 within the assembled dataset. However, the researchers did not test complete battery discharge performance, long-term cycling, or operation outside the measured chemical range.

The result pressures two familiar development strategies. One relies on testing many formulations and choosing a workable electrolyte from discrete experiments. The other treats the lowest predicted corrosion rate as the obvious answer. The Waterloo work argues that neither approach captures the full design problem because corrosion potential and corrosion current cannot always improve together.

The University of Waterloo Corrosion Model Changes Electrolyte Selection

The important change is not that AI predicted corrosion. It is that optimization converted a set of laboratory measurements into a specific, testable electrolyte condition.

Yasmin Shabeer, Seyed Saeed Madani, Satyam Panchal, and Michael Fowler studied a commercial aluminum-magnesium-silicon alloy. The work came from the University of Waterloo’s Department of Chemical Engineering. Funding included the Canada Research Chair program and the Natural Sciences and Engineering Research Council of Canada.

The researchers exposed alloy samples to sodium hydroxide, or NaOH, and KOH solutions. Each electrolyte was tested at concentrations from 1 to 10 molar. Tests ran at 25 degrees Celsius in a three-electrode cell, using aluminum alloy as the working electrode.

Every experiment began with 15 minutes of open-circuit stabilization. The team then used potentiodynamic polarization, which measures current while sweeping the applied voltage, to estimate corrosion behavior. Electrochemical impedance spectroscopy examined how the metal-electrolyte interface resisted a small alternating signal across different frequencies.

Two outputs anchored the analysis. Corrosion potential, written as Ecorr, indicates the electrochemical tendency associated with corrosion under the test conditions. Corrosion current density, or Icorr, serves as a comparative indicator of the corrosion reaction’s rate.

These measurements do not describe identical properties. A less negative corrosion potential can appear favorable while a rising current density signals faster material loss. That separation explains why choosing the best concentration from a single ranking can mislead researchers.

The experiments also found that electrolyte identity mattered. At matching concentrations, KOH generally produced a more negative corrosion potential and a higher corrosion current density than NaOH. The paper associates this difference with cation-dependent effects involving ionic mobility, hydration, and the electrical double layer near the metal surface.

Higher concentration did not produce a simple linear response. Low-to-moderate alkalinity promoted anodic activation, while high concentrations showed transport-limited and interfacial-limited behavior. In practical terms, adding more hydroxide eventually stopped producing proportional electrochemical changes.

X-ray diffraction detected no crystalline corrosion products after exposure. The authors interpret that result as continuous surface dissolution accompanied mainly by amorphous or short-lived surface films. It also leaves a gap because X-ray diffraction cannot fully describe corrosion shape, pit development, or nanoscale surface chemistry.

The researchers trained an artificial neural network on the electrochemical data. Its five inputs represented electrolyte type, concentration, temperature, applied potential, and measured current density. Its two targets were Ecorr and Icorr.

A grid search evaluated 351 neural-network configurations. The selected network contained three hidden layers with 20, 25, and 25 neurons. It used the Adam training algorithm, a learning rate of 0.001, and 500 training epochs.

The study reports R-squared values above 0.999 for both predicted outputs. R-squared measures how closely predictions track variation in the model’s reference data. Such a high value indicates a close fit within this dataset, but it does not establish accuracy across other alloys or operating environments.

That trained model became a fast surrogate for experiments. Rather than physically testing every intermediate concentration, optimization algorithms could query predicted corrosion behavior across a continuous search space. The process produced a proposed operating point between the original one-molar increments.

This is why the University of Waterloo corrosion model matters. It did more than reproduce measured results. It generated a narrower hypothesis that researchers can challenge with targeted experiments.

Why Aluminum-Air Batteries Need a Compromise, Not a Minimum

An aluminum-air electrolyte must help the anode release useful energy without consuming too much metal through parasitic corrosion.

In an aluminum-air cell, aluminum oxidizes at the anode while oxygen participates in the cathode reaction. The architecture draws one reactant from ambient air, helping reduce the mass carried inside the system. Aluminum also has a theoretical gravimetric energy density of 8.1 kilowatt-hours per kilogram as an active material.

That theoretical figure does not equal practical cell energy. Hardware, electrolyte, cathode components, incomplete utilization, and side reactions reduce system-level performance. Still, the underlying potential keeps aluminum-air chemistry relevant for applications where stored energy matters more than simple electrical recharging.

The electrolyte performs several jobs at once. It transports ions, supports aluminum dissolution during useful discharge, and influences reaction products at the anode. However, concentrated alkaline solutions can also attack aluminum while the cell is resting.

That unwanted reaction consumes active material without delivering useful current to an external load. It can also release hydrogen gas, creating efficiency, pressure-management, and safety concerns. The University of Waterloo paper therefore treats corrosion as a design variable, not a cosmetic materials issue.

This problem separates aluminum-air systems from conventional rechargeable lithium-ion batteries. In many aluminum-air designs, restoring capacity requires mechanical replacement or recycling of the consumed aluminum anode. Every gram lost through self-corrosion reduces the usable return from that replacement cycle.

Electrolyte concentration creates a particularly difficult choice. A weak solution can restrict ionic transport or useful reaction kinetics. A more aggressive solution can improve electrochemical activity while accelerating dissolution during operation or idle periods.

The Waterloo team formalized that conflict with two objectives. The optimization sought lower Icorr while moving Ecorr toward a less negative value. Improving one objective eventually degraded the other, creating a set of nondominated solutions called a Pareto front.

A Pareto front contains choices where no objective can improve without worsening another. The researchers selected its knee point, where further improvement in corrosion potential required a disproportionate increase in corrosion current. That point appeared near 1.35 M KOH.

The predicted values at that concentration were an Ecorr of negative 1.6199 volts and an Icorr of 0.0876 amperes per square centimeter. These numbers should be read together. Neither measurement alone explains why the algorithm preferred that condition.

Importantly, KOH was not the gentler electrolyte across the matched experiments. The study generally observed more aggressive corrosion in KOH than NaOH at equivalent molarity. Yet the optimizer still selected a KOH condition because it sought a particular balance across both objectives.

That apparent contradiction is the study’s most useful result. It shows why “lowest corrosion” is an incomplete design instruction. A condition that suppresses one indicator can produce an unfavorable electrochemical potential or constrain useful anode activity.

Prior research has already established corrosion as a bottleneck across multiple battery designs. A broad current-collector review describes how electrolyte composition, voltage, temperature, and aluminum surface properties affect dissolution in lithium-based cells. The component and mechanism differ from an aluminum-air anode, but the engineering lesson is similar: electrolyte benefits cannot be separated from metal compatibility.

For aluminum-air developers, the pressure is immediate. They need electrolyte formulations that support useful discharge, limit idle losses, manage hydrogen, and tolerate impurities. A model that narrows the search space can save experiments, but only if its objectives match the eventual battery’s requirements.

How AI Found the 1.35 M KOH Tradeoff

The model’s contribution comes from combining nonlinear prediction with a search method that preserves competing corrosion goals.

The artificial neural network learned relationships among the recorded electrochemical variables. Neural networks are useful when inputs and outputs interact nonlinearly, as they did across the concentration range. The model did not begin with a complete physical equation for every surface reaction.

The training process first normalized variables so values on different scales could be compared. The researchers then varied layer depth, neuron counts, learning rates, and an Adam optimizer parameter. This structured search identified a configuration that closely reproduced the available targets.

The reported training, testing, and validation R-squared values for Icorr were 0.99975, 0.99975, and 0.99970. A comparison expressed in physical units produced an R-squared value of 0.99974. These results suggest little visible error inside the study’s data split.

The next step used the model as an inexpensive predictor. A standard genetic algorithm searched for a single solution after combining the two objectives with equal weights. Genetic algorithms imitate selection, crossover, and mutation to improve candidate solutions over repeated generations.

Equal weighting introduces a judgment. It assumes corrosion potential and current density deserve equivalent influence after normalization. A different weighting scheme could favor another region of the design space.

The team therefore added NSGA-II, a multi-objective genetic algorithm that does not collapse both targets into one weighted score. It retained a range of nondominated electrolyte conditions and exposed the tradeoff between them. The knee point represented the candidate nearest an idealized combination of maximum Ecorr and minimum Icorr.

Both optimization approaches converged near 1.35 M KOH. The scalar method predicted an Ecorr close to negative 1.619876 volts and an Icorr of 0.087859 amperes per square centimeter. NSGA-II returned nearly the same location and measurements.

Agreement between two optimization strategies makes the selected point less dependent on one mathematical formulation. It does not provide independent experimental validation, however, because both methods queried the same trained neural network.

This difference matters when interpreting the University of Waterloo corrosion model. The convergence shows internal consistency across optimization methods. It does not show that a physical cell operated at 1.35 M KOH will outperform every alternative.

The approach still offers a practical shift in laboratory planning. Traditional screening would test concentrations at predefined intervals, compare measurements, and choose another set of experiments. A trained surrogate can explore intermediate values first, allowing researchers to spend laboratory time on the most informative candidates.

That strategy resembles active learning, where a model helps choose the next experiment instead of passively analyzing a completed dataset. The present paper did not report a repeating active-learning loop. Still, its workflow creates the foundation for one.

A future system could test the predicted optimum, add that result to the dataset, and retrain the model. Researchers could then sample areas where uncertainty remains high. Repeating that cycle would reveal whether the apparent knee point stays stable as evidence expands.

AI is increasingly used for related battery problems, including electrolyte-property prediction, materials screening, and interface analysis. A 2026 battery AI review highlights the same opportunities while identifying limited data quality, generalizability, and interpretability as persistent obstacles.

Those obstacles are especially important in corrosion. Surface preparation, alloy composition, dissolved impurities, flow, temperature, and test protocols can change measured behavior. A model can fit a clean experiment closely without capturing every variable encountered in a working system.

The Waterloo workflow is therefore better understood as experiment-guided optimization than autonomous chemical discovery. Humans defined the electrolyte families, concentration range, test procedure, target variables, and preferred tradeoff rule. The AI searched efficiently inside those boundaries.

For engineering teams, that division of labor is valuable. Models can preserve complex experimental relationships and rank follow-up tests. Researchers still decide whether the model’s objective represents battery value.

Near-Perfect Predictions Do Not Equal a Better Battery

The model’s exceptional fit is also the reason its boundaries deserve close attention.

An R-squared value above 0.999 sounds decisive. Yet prediction quality depends on the relationship between the training data and the question being asked. The study’s dataset came from one commercial aluminum-magnesium-silicon alloy, two hydroxide electrolytes, and one temperature.

The concentration range was broad, spanning 1 to 10 M. However, the underlying experimental grid remained structured and chemically narrow. The model was not asked to generalize across numerous alloy families, additives, contaminants, temperatures, or air-cathode designs.

Some inputs also require careful interpretation. Applied potential and measured current density were included among the five predictive features. Icorr was one of the outputs. Relationships among experimentally connected electrical variables can support accurate interpolation while offering less confidence under new operating regimes.

The paper says its data were divided into training, testing, and validation subsets. Random or closely related splits can place measurements from similar experimental conditions in every subset. That produces a demanding check on numerical fit, but not necessarily a harsh test of chemical extrapolation.

A stronger external test would hold back an entire concentration region, electrolyte family, alloy composition, or experimental campaign. Researchers could then examine whether the model predicts conditions it did not closely encounter during training. The published results do not establish that level of generalization.

The optimized 1.35 M concentration adds another verification requirement. The original experiments used solutions from 1 through 10 M, while the model searched continuously between measured points. A direct experiment at 1.35 M would determine whether the predicted knee survives physical testing.

The authors also state that direct correlations with aluminum-air discharge performance were outside the study’s scope. No full cell was used to prove higher delivered energy, improved shelf life, slower gas accumulation, or longer operating duration at the selected condition.

Corrosion potential and corrosion current density are valuable indicators, but a battery has more constraints. It needs adequate ionic conductivity, oxygen transport, cathode stability, discharge voltage, manageable reaction products, and acceptable behavior across its expected temperature range.

The test surface introduces another boundary. Samples were mechanically sanded, cleaned with ethanol, and exposed over 0.79 square centimeters. Commercial electrodes can have larger areas, different finishes, manufacturing residues, grain structures, and alloy impurities.

The team repeated electrochemical measurements at least three times and reported consistent trends. That supports reproducibility inside the chosen procedure. It does not remove scale-up uncertainty.

The use of continuous polarization sweeps creates another caveat. The researchers explain that Ecorr and Icorr were used for comparative analysis, rather than strict kinetic determination from ideal Tafel regions. Readers should therefore avoid treating the current values as universal lifetime predictions.

Surface evidence also remains incomplete. X-ray diffraction found no crystalline corrosion products, supporting the interpretation of continuous dissolution and amorphous films. The authors call for scanning electron microscopy or optical microscopy to clarify corrosion morphology.

That follow-up matters because equal average current densities can hide different failure patterns. Broad, uniform dissolution may affect an electrode differently from localized pits or cracks. A model trained only on summary electrical measurements cannot distinguish those outcomes without corresponding surface labels.

The optimization objective deserves scrutiny too. The knee point is a mathematical compromise between Ecorr and Icorr, not a universal definition of battery performance. A stationary backup unit, a range extender, and a disposable field power source can place different values on shelf life, peak output, and gas management.

Researchers might also prefer NaOH for reasons beyond the two modeled targets. The study observed lower corrosion aggressiveness in NaOH at equivalent concentrations. Material compatibility, supply considerations, freezing behavior, conductivity, and downstream recycling could alter a system-level choice.

Comparable battery research shows why broader validation is necessary. A benchmark investigation of aluminum current collectors found sharply different stability across five liquid electrolytes. Although that work addressed lithium-ion collectors, it illustrates how changing salts or solvents can overturn conclusions drawn from one chemical family.

Machine learning can reduce the number of experiments needed to locate promising conditions. It cannot make excluded variables irrelevant. Near-perfect interpolation should trigger more targeted physical testing, not less.

Three Tests Will Show Whether the Model Matters

The Waterloo result becomes consequential only if its predicted compromise survives direct experiments, broader chemistry, and full-cell operation.

The first signal is a physical test at approximately 1.35 M KOH. Researchers should repeat polarization and impedance measurements at that exact concentration, rather than relying on interpolation between integer molarities. Confirmation would strengthen the claim that the model found a real electrochemical knee.

That test should include neighboring concentrations. A narrow sweep around the proposed point would reveal whether 1.35 M marks a stable optimum or a mathematically precise answer on a shallow curve. Replication across separately prepared electrolyte batches would further test sensitivity.

Direct surface characterization should accompany the electrical measurements. Scanning electron microscopy can reveal localized damage, while chemical surface methods can examine oxide and hydroxide films. These observations would test the model’s implied transition between activation and transport-limited regimes.

If the measured Ecorr and Icorr match predictions, the University of Waterloo corrosion model gains credibility as an interpolation tool. A mismatch would weaken the specific optimum but still provide useful information for retraining.

The second signal is external validation across conditions the network did not see. Researchers could hold out a temperature, another aluminum alloy, or an additive-containing electrolyte. Success on those tests would matter more than another random split of closely related observations.

Temperature should receive early attention. The published experiments ran at 25 degrees Celsius, while practical systems may face colder starts and hotter operation. Reaction kinetics, conductivity, gas evolution, and film stability can all change with temperature.

Alloy variation is equally important. The tested material was approximately 99 percent aluminum with magnesium, silicon, iron, copper, chromium, titanium, and other minor constituents. Commercial anodes may use different alloying strategies specifically to control corrosion and discharge behavior.

Extending the chemical space would test whether the workflow generalizes beyond choosing between NaOH and KOH. Additives can suppress hydrogen evolution or modify surface films, but they also create more variables. That expanded space is where AI-assisted screening should offer its greatest advantage.

The third signal is full-cell performance. A useful aluminum-air electrolyte must deliver energy through a complete anode, separator, electrolyte, and air-cathode system. Researchers need to measure discharge voltage, aluminum utilization, hydrogen evolution, idle loss, and reaction-product management.

A full-cell test could also compare 1.35 M KOH with the strongest NaOH condition and a conventional baseline. That comparison would reveal whether the model’s two-objective optimum improves the outcomes that application developers actually buy.

Historical work shows why moving beyond half-cell measurements matters. Aluminum corrosion research has long found that electrolyte composition can alter pitting and protective-film behavior. A recent passivation study likewise connected aluminum stability to operating conditions in lithium-ion current collectors.

The chemistry and electrode roles differ, so those results cannot validate the Waterloo optimum. They do reinforce one lesson: a useful aluminum surface depends on the full electrochemical environment, not one isolated material property.

If all three signals align, the research could support a more efficient development loop. Experiments would generate structured corrosion data, a model would locate tradeoffs, and follow-up tests would correct the model. Full cells would then decide whether the optimized proxy variables create actual system value.

That workflow can also improve scientific record keeping. Battery teams must connect formulations, preparation steps, electrochemical traces, surface images, and model versions. A searchable engineering knowledge base can help preserve why one candidate advanced while another failed.

For now, the result is narrower but still useful. University of Waterloo researchers identified a nonlinear corrosion tradeoff and translated it into a falsifiable experimental target. They did not establish a commercial electrolyte, a longer-lasting cell, or a universal corrosion predictor.

The next move belongs in the laboratory. Test the 1.35 M KOH prediction directly, challenge the model with unfamiliar conditions, and compare complete cells against credible baselines. If those results agree, AI-assisted corrosion optimization will have moved from accurate interpolation toward practical battery design.