A battery materials campaign usually sets one demanding target, such as "keep 85% of capacity for 1,000 cycles while still delivering useful capacity at a ten-minute charge." Every cell that misses this target is filed as a failure. Yet each failed cell has still produced a full record of how it behaved: capacity at several charging speeds, and the shape of its fade over hundreds of cycles. Read honestly, that record says the cell succeeded at some easier target. This proposal takes a trick from automated theorem proving, hindsight relabeling, and applies it to the design of cathode materials. Hindsight relabeling means treating a failed attempt as a correct solution to whatever it did achieve. We relabel every cycled cell as a successful demonstration of the performance specification it actually met. A model that proposes materials, conditioned on a target specification, is then trained on these relabeled cells. It works its way toward the hard target through a curriculum of nearby, reachable targets.
The setting where this matters is multi-element doping of sodium-ion cathodes. Doping means substituting a small fraction of foreign metal atoms into the host crystal. Single dopants have been screened by hand. Combinations of two or three dopants at fine concentration steps run to thousands of recipes, and each recipe needs weeks of cycling before it can be judged. This is the regime where the combinatorial search problem and the long-cycling-stability problem meet. Two quantities must be paid for together: the number of candidate recipes, and the number of cycles needed to judge each one. A method that extracts more training signal per cycled cell attacks both.
Core Claim
The goal for this domain is a specification $g = (r, N, q)$: a cell meets $g$ if, cycled at current rate $r$, it keeps at least a fraction $q$ of its initial capacity after $N$ cycles. A cell trajectory $\tau$ includes a rate-step segment and a long-cycling segment. Its achieved set is
$$ A(\tau) = \{ (r, N, q) : C_r(N) / C_r(1) \ge q \}, $$
where $C_r(n)$ is the measured discharge capacity at rate $r$ on cycle $n$. The set is extended in the obvious way to capacity floors read from the rate-step segment. Our claim is that a goal-conditioned proposer $p_\theta(x \mid g)$ over recipes $x$ will reach a hard target specification in fewer cycled cells than two other learners given the same cells. The first comparison treats failures only as negatives to avoid. The second fits a surrogate to the cycling curves and ranks candidates with it. The proposer is a language model fine-tuned on pairs $(x, g)$ for every $g \in A(\tau)$. The first comparison is what the "error memory" of LLEMA does. The second is a curve surrogate in the style of BatteryLife's CyclePatch.
The early-cycle portion of $A(\tau)$ becomes available within days, while the headline target takes months. Relabeling therefore also produces training signal long before any cell finishes its final test.
Where This Sits
The mechanism comes from Proving Theorems using Incremental Learning and Hindsight Experience Replay. That system relabels failed first-order proof searches as proofs of the intermediate clauses they derived, which yields a self-paced curriculum of lemmas. On MPTP2078 it proved 1,424 theorems against 1,139 for E prover's auto mode. Its ablation lost 146 theorems when hindsight relabeling was removed. That ablation is our main evidence that relabeling, rather than the neural guidance alone, carries the gain. We reuse the idea but replace "intermediate clause" with "weaker performance specification." The goal space is then continuous and partially ordered, not a set of discrete logical formulas.
Three materials-search systems come closest to this work, and all three motivate it by what they leave out. Generative Hierarchical Materials Search prunes a language-to-crystal search with a formation-energy oracle. Matbench Discovery benchmarks stability screening against the zero-temperature convex hull, the set of most stable known compounds. LLEMA runs language-model evolutionary search with reward and error memories over static property targets. None of them sees a cycling curve.
LLEMA is the nearest neighbour in mechanism. Its error memory steers the search away from failed candidates, whereas we turn the same failures into positive examples for other specifications. A Synthesizability-Guided Pipeline for Materials Discovery acts as a constraint on our design. It realized only 7 of 16 top-ranked targets, and it covers only high-temperature solid-state routes. Our recipes therefore include sol-gel processing variables explicitly and are always physically made and tested, never only predicted.
Metastable Structure for Ultra-Sustainable, High Capacity and Kinetics-Enhanced Magnesium-Ion Battery adds a second constraint. Its best cathode is a phonon-unstable phase, one that is not the lowest-energy arrangement of its atoms, yet it cycled over 3,000 times at 1,000 mA g⁻¹. So we impose no convex-hull filter on proposals. The only filters are charge balance and the dopant-solubility limits observed in our own syntheses.
The application comes from Tailoring eg Orbital Occupancy of Fe in Ni-Doped Na4.3Fe3(PO4)2P2O7. It screened five single dopants (Ni, Mn, Co, Cu, Zn) in the sodium iron phosphate-pyrophosphate cathode NFPP, settling on Ni at x = 0.1, and left co-doping unexplored. In that study, eg orbital occupancy means how the outer d-electrons of iron fill one particular pair of energy levels. Doping shifts that filling, and the authors found a peaked, volcano-shaped relationship between it and the strength of the iron–oxygen bond.
Research Program
The search space crosses three choices: subsets of up to three dopants from {Ni, Mn, Co, Cu, Zn, Mg, Ti}, a total substitution level in six steps between 0.02 and 0.15, and three calcination temperatures. This gives 8,358 compositions before temperature, and roughly 25,000 recipes in total.
Every cell runs the same protocol, so that each trajectory spans many goals. The protocol is a rate ladder from 0.1C to 10C, followed by long cycling at 1C. Here 1C means a full charge or discharge in one hour, and 10C means six minutes.
The proposer is an open language model. It is first trained on the single-dopant literature recipes, then fine-tuned each round on relabeled pairs from all cells measured so far. To choose next-round goals, we take the frontier of $A$ achieved so far and step each goal toward the target $g^*$ along one axis at a time, either higher rate, more cycles, or higher retention. This mirrors the lemma curriculum in the theorem-proving work.
All arms in the study share one early-stopping rule, so that early stopping cannot explain any difference between them. A cell is retired once a fixed curve predictor, trained only on BatteryLife, gives it under 5% chance of meeting the current round's goals. The rule is frozen before the campaign starts. Its accuracy will be imperfect on a sodium cathode chemistry, since BatteryLife itself reports weak transfer across chemistries.
The target $g^*$ asks for more than the published Ni-doped reference. That reference reports 121.0 mAh g⁻¹ at 0.1C, 80.9 mAh g⁻¹ at 10C, and 89.1% retention after 1,000 cycles at 1C. We set $g^*$ as at least 90 mAh g⁻¹ at 10C together with at least 92% retention after 1,000 cycles at 1C.
Baselines
The naive reference points come in two kinds. The first is uniform random sampling from the recipe space under the same cell budget. The second is "do nothing," which simply re-synthesizes NFPP-Ni-0.1 in our own lab. Our own Ni-0.1 cells give the bar that any arm must clear, because the published figures above come from a different lab and setup.
The physics-descriptor baseline asks whether co-doping can be ranked without any learning. It takes the eg-occupancy volcano from the NFPP study, computes the occupancy of each candidate composition by density functional theory (DFT, the standard quantum-mechanical estimate of electronic structure), and tests the candidates closest to the volcano peak. This is the strongest prior a battery chemist would offer.
The LLEMA-style baseline is the same language-model proposer and the same evolutionary loop, but with reward and error memories in place of relabeling. Cells that meet the current goal enter the reward memory; all others enter the error memory as examples to avoid.
The strongest learned competitor is a surrogate-ranking arm. It uses the same proposer to generate candidates, then ranks them with a curve-predicting model trained on BatteryLife plus all of our measured cells, choosing by expected improvement toward $g^*$. The only reported number for this model class is a lifetime-prediction mean absolute percentage error of 0.179 on lithium-ion cells for CyclePatch (BatteryLife). No reported result was found on Lacuna for co-doped NFPP search, or for any cycling-aware materials search campaign.
Three ablations isolate the key ingredient. The first relabels only on the final retention value at the last cycle. The second drops the rate-ladder goals. The third replaces the curriculum with fixed sampling at $g^*$.
Fairness rules apply to all arms. Each arm gets an identical budget of four rounds of twelve recipes, with every recipe made as two replicate cells. All learned arms get the same base model, the same fine-tuning compute, and the same hyperparameter-search budget. All arms share the early-stopping rule. "Beating" a baseline means two things: reaching $g^*$ with fewer cells, and a higher best-found 10C capacity at matched 1,000-cycle retention, with a confidence interval from the replicate cells.
Why not just…?
Why not just fit a surrogate to whole cycling curves and use it for Bayesian optimization? This is the real threat. In principle a curve surrogate uses every failed cell's data just as relabeling does, so the information available to both is the same. The bet is narrower: a proposer trained on relabeled pairs generalizes better across a discrete recipe space of about 25,000 entries than a surrogate that can only rank what the proposer already offers, because the relabeled pairs teach the proposer directly what to generate. That is a hypothesis, not established fact. The surrogate-ranking arm shares the proposer and the data and differs only in how the data are used, so the comparison settles it. If that arm matches ours, we will report that relabeling adds nothing beyond curve regression.
Why not just use LLEMA's error memory? Error memory tells the search where not to go, but it discards what a failed cell shows about where it is possible to go. The theorem-proving ablation, which lost 146 theorems without relabeling, suggests this matters when successes are rare. Our ablation tests whether the same holds for cycling targets.
Why not just stop cells early using early-cycle lifetime prediction? Stopping cells early reduces the cost of each test. It does not decide which recipe to test next. We therefore give it to every arm equally rather than treating it as a competitor.
Why not just extend the eg-occupancy volcano to co-dopants? The volcano was fit to five single dopants and to capacity and band gap, not to long-term fade. The descriptor baseline measures directly how far this physics alone carries.
Proposed Experiments
The offline stage comes first, because it is cheap. On BatteryLife's 990 cells across 59 chemistries, we define recipes by chemistry and protocol fields, hold out the target specifications, and replay search as sequential selection from this pool, revealing each selected cell's curve. We compare cells needed to reach held-out targets across the relabeling, error-memory, and surrogate arms, using many random seeds. This pool is small and not combinatorial. It tests whether relabeling is sample-efficient, not whether it scales.
The wet-lab stage then runs the NFPP co-doping campaign described above, about 288 cells over four rounds. The checks are decided in advance. Relabeling must reach $g^*$ in fewer cells than the surrogate-ranking arm, or show a higher best 10C capacity at matched retention. Any recipe beating our own Ni-0.1 cells must reproduce in a fresh synthesis batch. At least one hit must hold up at lower electrolyte volume and higher electrode loading, closer to practical cells than standard flooded half-cell tests. If the offline gain appears but the lab gain does not, we will conclude that the bottleneck is synthesis noise rather than search, and the note will say so.
References
- LLEMA: Evolutionary Search with LLMs for Multi-Objective Materials Discovery
- BatteryLife: A Comprehensive Dataset and Benchmark for Battery Life Prediction
- Proving Theorems using Incremental Learning and Hindsight Experience Replay
- Generative Hierarchical Materials Search
- Matbench Discovery - Can machine learning identify stable crystals?
- A Synthesizability-Guided Pipeline for Materials Discovery
- Metastable Structure for Ultra‐Sustainable, High Capacity and Kinetics‐Enhanced Magnesium‐Ion Battery
- Tailoring eg Orbital Occupancy of Fe in Ni-Doped Na4.3Fe3(PO4)2P2O7 Cathode for High-Performance Sodium-Ion Batteries