Products Survive, Compounds Die: What 120 Invention Candidates Taught Us
The Numbers
We ran five rounds of a structured invention discovery loop — generate, critique, disprove, audit, learn, regenerate — evaluating 120+ candidates for commercially viable materials-science products. The cumulative kill rate was approximately 89%.
| Round | Candidates | Survivors | Kill Rate |
|---|---|---|---|
| R1 | 25 | 11 | 56% |
| R2 | 27 | 3 | 89% |
| R4 | 20 | 2 confirmed + 1 second-look | 85% |
| R5 | 48 | 0 clean (2 wounded) | 100% |
| Total | 120+ | ~16 | ~87% |
The Master Pattern: Products vs. Compounds
The single most important finding: service and API ideas survived at roughly 27%, while compound-level ideas survived at exactly 0%. All four compound promotions were killed because we chose well-published optima — the best-studied formulas in the literature are, by definition, already claimed.
Patentable compositions live in the gaps between published data points, not at the optima. Selecting the best-studied formula guarantees a prior art kill.
The Anti-Portfolio: 14 Named Kills
| Killed Idea | Kill Reason |
|---|---|
| Mixed-Electrolyte Space-Charge SSE | Unproven mechanism, single unreplicated study, multi-billion manufacturing gap |
| PGM-Free HER Catalysts by Durability | Pt cathode cost already low ($200-500/kW); 200h vs 80,000h durability gap; real bottleneck is OER not HER |
| Battery Degradation from Formation | Voltaiq (Siemens), Liminal, Monolith already dominate |
| Patent-Gated IP Portfolio | DFT-only patents unenforceable under 35 U.S.C. 112 (requires experimental data) |
| Electrolyte Formulation Optimizer | Aionics $10.6M, 4-year head start, CMU IP |
| Perovskite Degradation Oracle | No 25-year ground truth; 1000h→25yr extrapolation unreasonable |
| Thermoelectric ZT Screening | Alphabet Energy $52.7M deadpooled; prediction-to-experiment gap |
| Materials Dark Matter | GNoME already did this at Google scale; 380K structures = prior art |
| Coolant Compatibility Oracle | Decision-grade data is proprietary wet-lab; NIST TDE already reconciles what's public |
| Na-cathode composition | No FTO (Tiamat exclusive NVPF license); a composition is a paper, not a company |
| Patent-Clear LMFP Multi-Dopant Finder | Doped/fractional targets return 0 rows in warehouse; real fight is coating/process IP |
| SIB Patent-Thicket Heat Map | PatentsView feed deprecated (HTTP 410); incumbents sell it |
| PFAS-Exposure Liability Map | Crowded funded PFAS-compliance suites (IntegrityNext, 3E, Source Intelligence) |
| Spectroscopy-to-Structure Inversion | DiffractGPT free web app; Bruker/JEOL ship products; 10-100x overpriced |
Systemic Kill Patterns
- Named funded competitors are the sharpest kill. If someone with $10M+ and a 3-year head start already occupies the space, late entry is structurally dead.
- Wrong figure of merit kills instantly. PGM-free HER catalysts failed because the real bottleneck is OER, not HER — and Pt cathode cost is already below $500/kW.
- "Important" problems are crowded problems. PFAS and REE clusters scored 96-97 priority and produced 0/11 survivors.
- Prior deadpools in the space are strong negative signal. Alphabet Energy's $52.7M deadpool tells you more than any market analysis.
- A predictor whose validation data accrues at service-life speed is a lab business, not a data business. Perovskite degradation needs 25 years of ground truth. You don't have 25 years.
What We Changed
After these findings, we stopped generating compound-level candidates entirely and focused on service/API ideas where the defensible unit is the join — the cross-source data fusion — not the composition. We also added a mandatory prior-art pre-check before any compound promotion, and a forcing-function budget check before pursuing any high-priority domain.