\n\nProvenance: explore agent
\nagent-16(director-commissioned deep research, swarm of 7, 2026-07-21) — materialized verbatim, then editorially corrected only to replace the retracted 624/132 union-anchor figures with the reproducible primary-record values (558/154). Quantitative literature claims are as reported by the research agent from sources it accessed; see citations inline. Citation-verification pass pending before any external publication.
Scope note: per instructions, I did not re-survey the certified-inference per-step gating literature (coverage/selectivity theory, gate runtime costs). Everything below is economics: measured DFT-call budgets, labeling/abstention fractions, and accuracy-per-call pricing. Access level per source is marked FT (full text read this session) or ABS (abstract only). Evidence cut: 2026-07-21.
\n| Technique | \nWhat cost it removes | \nMeasured savings factor (system/size context) | \nAccuracy cost / failure mode | \nCitation (access) | \nEvidence strength | \n
|---|---|---|---|---|---|
| Bayesian-error gate, on-the-fly MLFF (VASP) | \nFP steps during FF training & generation | \n99% of FP calculations skipped during training, ~100x time cut even while learning; >99% bypassed for melting-point FFs of Al, Si, Ge, Sn, MgO | \nForce/energy error 2.6 meV/atom, 0.07 eV/Å vs FP; melting points reproduced quantitatively | \nJinnouchi, Karsai & Kresse, PRB 100, 014105 (2019), arXiv:1904.12961 ABS. Related earlier perovskite demonstration: Jinnouchi et al., PRL 122, 225701 (2019) FT | \nstrong | \n
| Model-deviation (max force σ over 4-model ensemble) gate, DP-GEN | \nDFT labeling of explored configurations | \n0.0044% of ~650M explored configs labeled (Al, Mg, Al-Mg, 50–2000 K) | \nUniform accuracy on defects/surfaces/phonons not in training data; failure mode = unlucky ensemble agreement in extrapolation | \nZhang, Lin, Wang, Car, E, PRM 3, 023804 (2019), arXiv:1810.11890v2 FT | \nstrong | \n
| QbC ensemble disagreement, chemical-space AL (ANI) | \nDFT data needed for a general organic-molecule potential | \nMatches prior potential with 10% of the data; beats it with 25% (COMP6 benchmark, CHNO) | \nResidual bias toward sampled regions; committee σ imperfectly correlated with error | \nSmith et al., JCP 148, 241733 (2018), arXiv:1801.09319 ABS | \nstrong | \n
| QbC + transfer learning to gold-standard QM (ANI-1ccx) | \nCCSD(T)/CBS evaluations | \nApproaches CCSD(T)/CBS on reaction thermochemistry/torsions; data-hungry in absolute terms | \nSmith et al., Nat. Commun. 10, 2903 (2019) FT (ChemRxiv preprint) | \nstrong | \n|
| SGP local-energy uncertainty gate (FLARE) | \nDFT calls during on-the-fly MD training | \n<100 DFT calls per 10 ps run across single- and multi-element systems | \nGP regression cost scales with sparse set; error controlled to ~meV/atom RMSE | \nVandermause et al., npj Comput. Mater. 6, 20 (2020), arXiv:1904.02042v3 ABS + FT snippet | \nstrong | \n
| SGP uncertainty gate, reactive FF (H/Pt) | \nDFT calls + human months for reactive FF construction | \n250 DFT calls total (216 in 3.7 ps reactive run); 65.5 h wall; then 500 ps production MD; >10x post-mapping speedup, 2x faster than ReaxFF | \nActivation energy 0.25(2) eV vs exp 0.21–0.24; C44 within 6% (ReaxFF off by 200%) | \nVandermause et al., npj Comput. Mater. (2022), PMC9440250 FT | \nstrong | \n
| Uncertainty-biased MD sampler (UDD-AL) | \nMD steps (hence labels) to reach high-uncertainty/reactive regions | \n350 K biased run matches 600 K unbiased coverage with far fewer steps; 90 proton transfers/0.5 ns at 350 K vs 0 unbiased (48 at 620 K); glycine dataset built with 1,405 DFT labels, RMSE <0.3 kcal/mol vs 50k-structure reference | \nTwo empirical bias parameters (A, B); not automated | \nKulichenko et al., Nat. Comput. Sci. 3, 230–239 (2023), PMC10766548 FT | \nstrong | \n
| Uncertainty-biased HAL sampler (ACE) | \nDatabase assembly cost for alloy/polymer potentials | \nACE potential for AlSi10 from 88 DFT configurations (32 atoms each), starting from ~dozen; melting T and density near experiment | \nBias strength τ heuristic; small-cell bias | \nvan der Oord et al., npj Comput. Mater. 9, 168 (2023), arXiv:2210.04225 ABS + FT snippet | \nmoderate | \n
| Extrapolation-grade (maxvol) gate, MTP for alloy convex hulls | \nDFT relaxations in high-throughput screening | \n3–4 orders of magnitude faster than conventional HT-DFT; found unreported stable structures vs AFLOW (Cu-Pd, Co-Nb-V, Al-Ni-Ti) | \nFactor folds in surrogate speed, not only call reduction; grade misses smooth-interpolation errors | \nGubaev et al., Comput. Mater. Sci. 156, 148 (2019), arXiv:1806.10567v1 ABS | \nmoderate | \n
| Δ-ML correction on cheap baseline | \nHigh-level QM labels for large molecular sets | \nSemiempirical+ML trained on 1–10% of 134k molecules reproduces DFT-level enthalpies of the remainder (16k C₇H₁₀O₂ isomers to chemical accuracy) | \nOnly static properties; baseline must capture physics | \nRamakrishnan et al., JCTC (2015), arXiv:1503.04987v1 ABS | \nstrong | \n
| Δ-ML + rank-compression gate for beyond-DFT MLFF (zirconia) | \nRPA/hybrid-level calculations | \n168 RPA structures (24-atom cells) selected from 1,275; <150,000 CPU hours total RPA training; could halve further | \nPhase-transition T within ~50 K of experiment; needs DFT-level on-the-fly pass first (592 structures, 96-atom) | \nLiu, Verdi, Karsai, Kresse, PRB 105, L060102 (2022) FT | \nstrong | \n
| GP-surrogate gate in NEB (GP-NEB) | \nAb initio energy/force evaluations per MEP | \nOrder-of-magnitude fewer evaluations vs conventional NEB (benchmark: island rearrangement transitions) | \nGP struggles with short-distance large-force configs (fixed in 2019 inverse-distance revision, JCTC 15, 6738; dimer variant JCTC 16, 499 (2020)) | \nKoistinen et al., JCP 147, 152720 (2017), arXiv:1706.04606 ABS | \nstrong | \n
| On-the-fly GPR surrogate inside NEB (GPR_calculator) | \nElectronic-structure calls in massive NEB campaigns | \n3–10x acceleration vs pure ab initio NEB (surface diffusion/reaction demos) | \nPreprint, small demo set; no failure-rate audit | \nZhu et al., arXiv:2504.07319v2 (2025) ABS | \nclaim | \n
| ML-rank + top-k DFT verify (AdsorbML) | \nDFT relaxations in adsorption-energy screening | \n87.36% success at ~2000x speedup; balanced option 86.33% at 1331x (OC20-Dense, ~1000 surfaces, ~100k configs) | \n~13% of systems fail to recover the DFT minimum within 0.1 eV; the verify step is the residual DFT cost | \nLan et al., npj Comput. Mater. 9, 172 (2023), arXiv:2211.16486v3 ABS + FT snippet | \nstrong | \n
| Ensemble + optimization acquisition to steer DFT (Tran & Ulissi) | \nDFT evaluations in catalyst discovery | \nScreened alloys of 31 elements (50% of d-block, 33% of p-block); 131 CO₂R candidate surfaces/54 alloys; 258 HER surfaces/102 alloys | \nExact call-reduction factor [UNVERIFIED] — full text paywalled; abstract gives scope, not the factor | \nTran & Ulissi, Nat. Catal. 1, 696–703 (2018) ABS | \nmoderate | \n
| Conformal-calibrated latent-distance UQ (CP+latent) | \nRe-training cost of ensemble/dropout UQ; false alarms | \nCalibrated and sharp intervals; applied to 1M-image training set at low cost (OC20-family NNFFs) | \nGuarantee void under i.i.d. violation (their own tests); no DFT-call budget reported — savings implied, not priced | \nHu, Musielewicz, Ulissi, Medford, MLST 3, 045028 (2022), arXiv:2208.08337v2 ABS | \nmoderate | \n
| Statistical error-cutoff for gate thresholds | \nOracle-error information in threshold setting; cost of ensemble UQ | \nCutoff-gated AL with cheap UQ (sparse GP, latent distance) matches true-error-oracle AL on 2 datasets × 3 UQ methods | \nDocuments that raw ensemble σ decorrelates from true error — gate on cutoff, not σ | \nAnnevelink & Viswanathan, arXiv:2308.15653v1 (2023) ABS | \nmoderate | \n
| Calibrated gradient-based (single-model) UQ + uncertainty-biased MD (UBMD) | \nEnsemble compute in AL sampling | \nEnsemble-comparable or better MLIPs at lower computational cost (alanine dipeptide, MIL-53(Al)) | \nCalibration needed; bias parameters manual | \nZaverkin et al., npj Comput. Mater. 10, 83 (2024), arXiv:2312.01416v2 ABS | \nmoderate | \n
| Evidential single-pass UQ gate (eIP / e²IP) | \nEnsemble forward passes in AL/UDD and production monitoring | \nReliable UQ "without significant computational overhead"; AL + UDD demos; universal potential with real-time UQ (eIP); equivariant force-covariance version claims better accuracy-efficiency-reliability balance than ensembles (e²IP) | \nYoung, single-lineage claims; no independent replication of gate-selectivity numbers | \nXu et al., arXiv:2407.13994v2 (2024) ABS; Wang et al., arXiv:2602.10419v2 (2026) ABS | \nclaim | \n
| CP-inspired learnable quantile calibration on a foundation model | \nSharpening of uMLIP intervals before gating | \nOrder-of-magnitude better uncertainty-error correlation; improved AL data efficiency; cross-functional transfer of calibration (MACE-MP-0) | \nPreprint; "negligible cost" unquantified in DFT-call terms | \nHo et al., arXiv:2510.00721v1 (2025) ABS | \nclaim | \n
| Distance-to-training-set adaptive gate in AIMD (Botu & Ramprasad) | \nFP steps in AIMD | \nAdaptive on-the-fly learning of forces; no factor in abstract [UNVERIFIED factor] | \nEarly descriptor/kernels; superseded by GP/NN gates | \nBotu & Ramprasad, IJQC 115, 1074 (2015) ABS snippet | \nclaim (historical) | \n
| Two-member committee NNP gate (Schran/Marsalek) | \nEnsemble size for disagreement UQ; generalization-error control | \nCommittee of 2 NNPs bounds generalization error and drives AL (bulk water, aqueous systems) | \nNumber of labeled configs not abstract-quantified [UNVERIFIED budget] | \nSchran, Brezina, Marsalek, JCP 153, 104105 (2020) citation + secondary | \nmoderate | \n
Replicated and trusted (multiple independent groups, numbers reproduced):
\nSingle-group or young claims (real numbers, not yet replicated):
\nMarketing-adjacent (true but inflated framing):
\nNotably thin (honest assessment): true production-MD abstention rates (post-training, the fraction of deployed inference sent back to DFT) are almost never reported; nearly all published fractions are from training campaigns. AdsorbML's ~13% verify tail and the on-the-fly "calls become infrequent after 3–10 ps" observation (FLARE) are the only production-mode numbers I could verify. Conformal/evidential gates likewise have no published DFT-call budgets — they sell calibration, not priced abstention.
\n[UNVERIFIED items flagged inline: Tran & Ulissi's exact DFT-call factor (paywalled full text); Botu & Ramprasad's speedup factor; Schran committee's labeling budget; Hodapp & Shapeev "in operando active learning" (MLST 1, 045005, 2020 — citation verified, numbers not accessed). All other quantitative claims above were read this session at the access level stated.]
\n"}