Transfer Learning from Lithium-Ion to Zinc-Ion Batteries: Cross-Chemistry Capacity Prediction with Limited Target Data

Zinc-ion batteries offer real advantages for stationary storage—zinc is abundant, aqueous electrolytes are inherently safe, and manufacturing costs remain low. Yet compared with lithium-ion systems, the published literature on zinc-ion cathode materials remains thin. Most machine-learning efforts in battery research have focused on lithium chemistries, where thousands of experimentally characterized compounds are available. For zinc-ion cathodes—manganese oxides, vanadium bronzes, Prussian blue analogues, organics—I could locate only about 75 usable entries with consistent reporting of initial discharge capacity, voltage window, and current density. This limits predictive modeling: standard regression models overfit or underperform on such small sets. Can knowledge encoded in larger lithium-ion datasets transfer to zinc-ion prediction tasks, or does the cross-chemistry gap introduce more bias than benefit? To test this, I built a cross-chemistry feature schema containing seven variables shared between both battery types: minimum and maximum cutoff voltage, average operating voltage, voltage window, current density as a rate proxy, a crystal-structure encoding, and a dimensionality encoding. These descriptors deliberately exclude chemistry-specific markers like ionic radius or intercalation mechanism, focusing on operational and structural parameters common to both systems. The lithium-ion source domain contains 150 synthetic entries sampled from distributions reported in published studies across five cathode families (LCO, LFP, NMC, NCA, LMO). The zinc-ion target set has 75 records from the experimental literature covering manganese oxides, vanadium oxides, Prussian blue analogues, and organic cathodes. I compared three model architectures—Random Forest (RF, 200 trees), Gradient Boosting Regression (GBR, 200 estimators, learning rate 0.1), and a multilayer perceptron (MLP, 64-32-16 ReLU)—across five training strategies under 5-fold cross-validation: zinc-only fitting, naive combined pooling, instance-weighted transfer (GBR, source weight w=0.3), MLP pretraining on lithium-ion data followed by fine-tuning, and a few-shot learning curve that tracks performance as the zinc-ion fraction rises from 20% to 100%. With all 75 zinc-ion records available, the zinc-only RF achieved the best performance: R²=0.473±0.172, MAE=47.2 mAh/g. The combined RF, which pooled both datasets without adaptation, reached R²=0.463±0.208 and MAE=46.7—marginally better on absolute error but with higher variance. GBR showed a similar pattern: zinc-only gave R²=0.350±0.291 (MAE=48.8), combined reached R²=0.416±0.295 (MAE=47.4), and instance weighting yielded R²=0.384±0.301 (MAE=49.8). Neither MLP setting was usable. Zinc-only MLP produced R²=-0.910±1.854 (MAE=91.1), and the transfer variant was worse at R²=-1.228±0.753 (MAE=111.0). With roughly 60 training samples per fold, the network cannot learn stable representations, and lithium-ion pretraining pushes it into a region mismatched with zinc-ion electrochemistry—negative transfer caused by mechanistic differences: aqueous Zn²⁺ insertion involves solvation shells and proton co-insertion absent in organic-electrolyte Li systems, and some Zn cathodes undergo partial conversion reactions with no lithium-ion counterpart. The few-shot analysis clarifies where transfer pays off. At 20% of the zinc-ion data (roughly 15 samples), zinc-only GBR collapsed to R²=-0.555, while the transfer approach recovered to R²=-0.209. At 80%, transfer held an edge: R²=0.361 versus 0.266. The benefit concentrates in the smallest splits; as zinc-ion labels accumulate, the gap compresses. For experimentalists working with new cathode families, this low-data regime is often what matters most—even a modest accuracy gain at 12 to 30 samples can make a model useful for ranking candidate materials before committing to extensive cycling campaigns. Feature importance from the weighted GBR confirmed an electrochemically sensible picture: current density dominated (0.582), followed by average voltage (0.389) and crystal-structure encoding (0.350). Rate and voltage constrain accessible capacity in aqueous systems; the structure code captures diffusion-pathway differences between oxide and organic host frameworks. Cross-chemistry transfer from lithium-ion to zinc-ion data provides conditional gains rather than a universal improvement. If a moderate zinc-ion dataset is available, a zinc-only RF baseline is hard to beat. If zinc-ion data are scarce, pooling or weighting lithium-ion records can stabilize early-stage predictions—but only with tree-based methods. Neural transfer is not viable at this sample size. Collecting chemistry-specific experimental data remains essential for accurate capacity prediction; lithium-ion data can inform early models but cannot replace zinc-ion experiments. Figure 1