A Catalog of Real Failures

Over the course of building a 19 GB materials warehouse aggregating data from 9+ DFT databases and running our own compute campaigns, we encountered a comprehensive set of failure modes. These are not hypothetical risks — every one of these was discovered in production data that fed our invention discovery pipeline.

Compute Failures

1. Thermostat Overshoot

Li2HfO3: Target temperature 625K, actual simulation temperature 750K — a 20% overshoot. AIMD stability claims at 625K are invalidated when the thermostat actually ran at 750K. The result may still demonstrate stability, but at the wrong temperature.

2. MACE Artifacts on Perovskites

BaZrO3: The MACE universal potential predicted an Imma space group that doesn't match any known BaZrO3 phase. This is a force-field artifact, not a physical prediction. MACE overstiffness on perovskite structures produces fictitious phase assignments.

3. PBE Bandgaps Without HSE Correction

AgPS3: Compute results show PBE extraction, but the pipeline presented them as-is without flagging that PBE systematically underestimates bandgaps by 30-50%. Filing a patent claim about a bandgap range without HSE06 correction is risky.

4. Metallic DFPT Results

FeBiO3: Band gap extracted as zero or negative. Running DFPT on a material that DFT predicts as metallic is physically meaningless — the Born effective charges and dielectric tensor have no valid interpretation for a metal.

5. Error Payload Masquerade

All 13 Modal QE results from April 2026 were error payloads marked as DONE. Files downloaded successfully, had proper filenames, and passed existence checks — but contained only error messages. Without status-code validation, these would have been ingested as real data.

Data Interpretation Errors

6. THz vs cm⁻¹ Unit Confusion

53 files affected. Phonon frequencies reported in THz in some sources and cm⁻¹ in others, with no unit metadata. 1 THz ≈ 33.356 cm⁻¹ — misinterpreting the unit by ignoring this factor makes a stable material look unstable (imaginary frequencies) or vice versa.

7. Alias Confusion

BiRhSe vs BiIrSe: Rh and Ir are in the same column of the periodic table. A compute result tagged as BiRhSe was actually computed for BiIrSe (same structure file, wrong label). This class of error — same crystal structure, wrong element label — is nearly undetectable without checking the pseudopotential files used.

8. Phonon Instability Misinterpretation

Rb2SnO3: Phonon instability at -0.61 THz flagged as a ferroelectric soft mode. But the extraction confirmed it is NOT a ferroelectric soft mode — it's a genuine structural instability. Confusing dynamic instability with soft-mode-driven ferroelectricity leads to false claims about ferroelectric properties.

9. Anisotropy Averaging

Na2PdS2: Dielectric anisotropy ratio of 2.28× averaged into ε_inf_mean=7.34. Reporting only the mean obscures the fact that one crystallographic direction has dramatically different properties — relevant for any application that depends on directional response.

Process and Pipeline Errors

10. Haiku Hallucination

Using Claude Haiku for candidate evaluation produced confident-sounding assessments of capabilities the model cannot deliver — specifically, MACE-POLAR-1 capability overestimation. Four candidates were built on capabilities the potential does not actually have.

11. Stale Snapshot Drift

Cached warehouse snapshots used for screening diverged from the live warehouse by 12.6M rows. Screening decisions made against stale data are no better than guesses when the underlying data has materially changed.

12. Sub-Agent Context Leak

When using multi-agent AI workflows for invention discovery, context from one evaluation leaked into another, causing one candidate's assessment to be influenced by prior candidates in the queue. Evaluation order should not change verdicts.

13. Ceder Dataset Thesis Invalidation

An invention thesis was built on the assumption that Ceder synthesis recipes covered a specific compositional space. The Ceder group's own dataset update invalidated this premise — the thesis was sound against v1 of the data but false against v2.

Defensive Recommendations

• Validate compute status codes before reading result fields
• Carry units metadata through the entire pipeline, never infer
• Check pseudopotential filenames against reported compositions
• Report per-direction properties, not just isotropic averages
• Timestamp every cached artifact and invalidate on warehouse refresh
• Isolate multi-agent evaluation contexts