This page reports the public benchmark scope for FluxMateria's battery-native materials layer: calibrated holdout accuracy, scenario alignment across distinct battery questions, and the end-to-end workflow outcome from the latest local battery study.
Two benchmark layers were used because a serious battery engine should be accurate and also change its answer when the engineering question changes.
The battery layer was evaluated against known cathode-family references using calibration rows plus an untouched holdout slice.
The same layer was then tested across different battery decision frames instead of only a single blended score.
Important scope note: this benchmark validates FluxMateria's battery-native decision layer as a screening, triage, and prototype-handoff engine. It does not claim complete electrochemical or wet-lab validation. The point is to reduce the search space to better-supported build candidates and make the next physical experiment sharper.
The holdout result checks whether the battery layer is directionally and quantitatively aligned with known battery families before it is used for novel ranking.
| Metric | Result | Interpretation |
|---|---|---|
| Family accuracy | 1.0 | Every holdout reference was assigned to the correct high-level battery family. |
| Capacity MAE | 0.812 mAh/g | Specific-capacity estimates stayed close to nominal literature-aligned reference values. |
| Voltage MAE | 0.149 V | Average-voltage prediction remained within a tight screening-grade range on holdout materials. |
| Transport MAE | 0.1427 | Transport and rate-readiness signals were directionally consistent with known family behavior. |
| Cycle-life MAE | 0.0642 | Cycle/degradation heuristics stayed close to the nominal reference scoring used for calibration. |
| Electrolyte MAE | 0.09 | Electrolyte compatibility stayed well aligned with the known chemistry tradeoffs in the holdout slice. |
| Interface MAE | 0.0883 | Interface-readiness scoring tracked the reference set closely enough for shortlist triage. |
| Cost / manufacturing MAE | 0.0372 / 0.0765 | Cost and practical build signals stayed stable enough to support the handoff layer. |
| Energy-rank Spearman | 0.9429 | The model preserved the energy-ordering structure of the holdout references. |
The battery layer was then stressed across different engineering questions. All five public scenarios aligned with the intended family or material outcome.
| Scenario | Primary metric | Observed top result | Margin | Why it matters |
|---|---|---|---|---|
| Energy-dense cobalt-free screen | battery_readiness_score | LiMnO2 | 5.2 | The energy-focused screen lifted layered manganese oxide instead of collapsing back to cobalt-heavy chemistry. |
| High-voltage frontier | voltage_surrogate_V | LiNiPO4 | 0.2 V | The voltage layer correctly pushed very-high-voltage phosphate chemistry to the top when voltage itself was the question. |
| Fast-charge transport | rate_capability_proxy | LiMn2O4 | 0.006 | The screen surfaced a 3D spinel transport leader, with Li4Ti5O12 essentially tied as the other valid 3D transport winner. |
| Cycle-life conservative | cycle_life_proxy | Li4Ti5O12 | 0.124 | The long-life framing favored the most stability-oriented chemistry rather than the highest-energy candidate. |
| Immediate build handoff | prototype_handoff_priority_score | Li4Ti5O12 | 4.7 | The handoff layer separated the best immediate prototype package from the highest-upside chemistry. |
The same battery workflow produced different leaders as more battery-native logic was added. That is a feature, not a bug.
What this means: FluxMateria is not behaving like a single-score materials ranker. The battery layer changes its answer when the engineering question changes. That is exactly what a usable battery decision engine should do. Bulk energy density, interface readiness, balanced electrochemistry, and immediate prototyping are related questions, but they are not the same question.
Public benchmark materials for independent review and reuse.
Review the full battery case study, then see how the battery layer fits inside the broader FluxMateria materials stack.