{"id":"savings-active-learning","title":"Active Learning for Interatomic Potentials as a DFT-Call Economizer","subtitle":"Active-learning loops for training interatomic potentials, priced in DFT calls avoided.","category":"references","tags":["literature-review","savings-stack","active-learning","mlip"],"source":"articles/lit-review/savings-active-learning.md","lang":"en","words":2304,"readMinutes":10,"toc":[{"depth":2,"text":"1. Executive result","id":"1-executive-result"},{"depth":2,"text":"2. Savings table","id":"2-savings-table"},{"depth":2,"text":"3. Proven vs claimed","id":"3-proven-vs-claimed"},{"depth":2,"text":"4. Openings for Lupine","id":"4-openings-for-lupine"}],"html":"
\n

Provenance: explore agent agent-14 (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.

\n
\n

Chapter Digest: Active Learning for Interatomic Potentials as a DFT-Call Economizer

Scope note. This subdomain is not thin — it is one of the most densely evidenced compute-savings areas in atomistic simulation. Every number below was read this session from the cited arXiv abstract page or full text (marked accordingly); journal-only items were cross-checked via Crossref/ACS metadata. Evidence cut: 2026-07-21.

\n

1. Executive result

\n

2. Savings table

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
TechniqueWhat cost it removesMeasured savings factor (system/size context)Accuracy cost / failure modeCitation (what I read)Evidence strength
VASP MLFF on-the-fly Bayesian AL (2-/3-body + SOAP-like descriptors, Bayesian error on forces)DFT calls during training MD; AIMD for production runs>99% of FP calculations bypassed during FF generation; ~1000× simulation speedup; melting points of Al, Si, Ge, Sn, MgOOnly samples configurations the trajectory visits; validation vs thermodynamic perturbation neededarXiv:1904.12961v3, abstractstrong
VASP MLFF perspective (scaling of the above)Same, generalized to large-scale MD"most FP calculations bypassed… accelerated by several orders of magnitude"Qualitative; no per-system accountingJ. Phys. Chem. Lett. 11, 6946 (2020), DOI 10.1021/acs.jpclett.0c01061, abstractstrong
Learn-on-the-fly (Li–Kermode–De Vita), GP extrapolation gradePeriodic re-evaluation of QM during MDFactor ~30 fewer QM calculations (as reported in the Podryabinkin–Shapeev full text)Fixed-interval fallback needed; decision not geometry-drivenPRL 114, 096405 (2015), quoted in arXiv:1611.09346v3 full textmoderate (number verified as quoted, original not read)
MTP D-optimality active learning (MaxVol, extrapolation grade γ)QM calls in MD/relaxation; guarantees interpolation-only predictionOn-the-fly training with QM calls "typically in the initial stage"; no extrapolation attempted; threshold γ_th tunableRestricted to linear-in-parameter models; γ is geometric, not an error barPodryabinkin & Shapeev, CMS 140, 171 (2017), arXiv:1611.09346v3, abstract+full textstrong
ANI query-by-committee (ensemble disagreement)Data needed for transferable organic-molecule NNPParity with ANI-1 on COMP6 at 10% of data; dominant at 25% (~4–10× vs naive sampling)QBC blind where committee shares bias; per-molecule DFT cost still ~10⁵–10⁶ points for universal setsarXiv:1801.09319v2, abstractstrong
DP-GEN concurrent learning (model deviation of 4-model DP ensemble)Labels for broad (T,p) coverage of alloysAl: 6,489 / Mg: 3,689 / Al–Mg: 18,654 labels from ~650M explored (0.0044%); uniform accuracy 50–2000 K incl. defects, surfaces, liquidConfigs with deviation > σ_hi (0.15 eV/Å) discarded as "unphysical" — rare high-error states can be dropped unlabeled; no rigorous indicator theory (authors' own caveat)arXiv:1810.11890v2, full textstrong
DP-GEN platform (scheduler/dispatcher)Human labor + labels for general-purpose potentialsCu: 7,646 labels (0.03%) from 25M explored, 48 iterations, 0–2T_m, 1 bar–50 kbarSame trust-band caveat (σ_lo 0.05, σ_hi 0.20 eV/Å); labeling caps at 300 configs/iterationarXiv:1910.12690v1, full textstrong
Committee NNP (c-NNP, committee disagreement monitored/biased in MD)Ab initio set for water incl. NQE814 reference calculations total for liquid water (ambient + elevated T,p), ice phases, air–water interface, with path-integral NQECommittee = multiple models → training cost multiplied; state-point-specific, not universalarXiv:2006.01541v2, abstractstrong
Committee AL for aqueous interfaces (CP2K/NNP)Model construction effort for complex aqueous systemsAutomated AL after one initial AIMD run; validated structure/dynamics for ions, TiO2/water, confined waterCommittee-size and validation protocol required; single-state-point modelsSchran et al., PNAS 118, e2110077118 (2021), arXiv:2106.00048v2, abstractstrong
FLARE (low-dim GP, Bayesian epistemic/noise uncertainty)Training data for on-the-fly force fields; AIMD cost for rare-event MD~100 DFT calls/system (C:107, Si:133, Al₂O₃:16, NiTi:18, BN:237, AgI:39); Al vacancy 1 ns training run >300× faster than equivalent AIMD; HEA: GP matched NNs trained on >10× more structuresGP epistemic uncertainty has an upper bound — true error underestimated for δ>20% distortions; threshold tied to noise σ_n needs care at phase changesarXiv:1904.02042v3, full textstrong
FLARE++ (sparse GP + ACE descriptors, mapped to polynomial)Training data + inference cost for reactive force fields216 DFT calls (reactive Pt/H) + 24/4/6 (H₂, Pt surface, bulk) ≈ 250 total → reactive FF in 3 days; 2× faster than Pt/H ReaxFF; E_a 0.25(2) eV vs exp 0.23 eVSingle-system demonstration; extrapolation confidence grows outside training set but flagged only by GP varianceVandermause et al., Nat. Commun. 13, 5183 (2022), arXiv:2106.01949v1, full textstrong
GAP iterative learning (a-C)Full AIMD melt–quench trajectories (~9,500 steps) per uncorrelated amorphous sampleProtocol: preliminary GAP drives melt–quench; only 1+ DFT single-point per GAP trajectory retained; iteratively extended across densities 1.5–3.5 g/cm³Early many-body-only model produced sub-Å atom aggregation (unphysical extrapolation); total DFT count not stated in main text [UNVERIFIED — given only in SI]Deringer & Csányi, PRB 95, 094203 (2017), arXiv:1611.03277v1, full textstrong (protocol), [UNVERIFIED] total count
ACE D-optimality vs ensemble (MaxVol extrapolation grade)Ensemble training cost for ACE; rare-event discoveryD-optimality indicator "more computationally efficient" than ensembles at comparable prediction; enables automated rare-event exploration incl. per-environment selection in large MDTwo query strategies compared, no single hard DFT-call factor reportedLysogorskiy et al., PRM 7, 043801 (2023), arXiv:2212.08716v1, abstractstrong
MLIP-3 per-atomic-environment AL (MTP)Whole-configuration labeling in large simulationsActive learning on atomic neighborhoods, enabling AL in large-scale runs where whole-frame selection is too coarseSoftware/methods paper; efficiency claims qualitativePodryabinkin et al., JCP 159, 114104 (2023), arXiv:2304.13144v3, abstractstrong
FALCON OTF calculator (AGOX GPR, ASE)DFT calls in long MD; retraining overhead via k-means clustered sub-modelsAl₃₂ melting, 250 ps: 500 DFT calls in first 25 ps, then only 69 in remaining 225 ps; H₂O-in-CNT (115 atoms): 151 DFT calls for 50,000 steps, >300× vs AIMD (500 days → 38.4 h)DFT-call count swings ~4× with threshold choice (0.10→0.20 eV: >300→86 trainings); GPR retraining can exceed AIMD cost for large/hot/disordered systems (authors' own warning)Felis & Dononelli, npj Comput. Mater. 12 (2026), DOI 10.1038/s41524-025-01897-8, full textstrong
Small-cell AL (1–8 atom training cells)Large-cell DFT cost during ALUp to 100× cost savings vs 54-atom-cell training; some runs <120 core-hours (K, Na–K; solid–liquid interfaces, critical exponents reproduced)Single group; two alkali-metal systems; generality unprovenMeng et al., Comput. Mater. Sci. 256, 113919 (2025), arXiv:2504.07293v1, abstractmoderate→strong
Uncertainty-biased MD (energy-uncertainty bias forces + bias stress)Missed rare events / extrapolative regions of plain AL-MDMLIPs "similar or better accuracy than ensemble-based methods at lower computational cost" (alanine dipeptide, MIL-53(Al))Requires calibrated gradient uncertainties; new method, limited replicationarXiv:2312.01416v2, abstractmoderate
On-the-fly fine-tuning of foundation models (Bayesian NN UQ)Training-data burden when specializing pretrained uMLIPsAutomates fine-tuning at pre-specified accuracy; "detects rare events such as transition states and samples them at an increased rate"No hard DFT-call factor in abstract; foundation models lack native UQRensmeyer et al., Digital Discovery (2026), DOI 10.1039/D5DD00392J, abstractmoderate
LOCAL (GNN + locality AL, catalyst screening)DFT labels for stability screening of 611,648 DAC/NG structures16,704 labels (2.7%) → 0.15 eV test MAEScreening task (not MD potential); single paperYin et al., arXiv:2503.19445v5, abstractmoderate

3. Proven vs claimed

Replicated and trusted (multi-group, multi-codebase):

\n\n

Single-paper or thin-replication claims:

\n\n

Marketing / handle-with-care:

\n\n

4. Openings for Lupine

    \n
  1. Theorem-gating fills the documented "confident-but-wrong" blind spot of every learned-UQ trigger. DP-GEN's authors concede the committee indicator is sufficient-but-not-necessary (ensemble can agree and be wrong, arXiv:1810.11890 full text); FLARE's GP variance saturates and underestimates error under strong extrapolation (arXiv:1904.02042 full text); Tan et al. benchmarked that no UQ scheme consistently flags OOD (npj Comput. Mater. 2023, arXiv:2305.01754v1, abstract). A Lean-formalized physical-law check (symmetry, conservation, curvature-sign at saddles) is an independent, non-learned oracle that can veto predictions exactly where committees and GP variances are structurally blind — Lupine should frame theorem-gating as the necessary-condition complement to AL's sufficient-condition triggers.
    2. \n
  2. Union anchors are an orthogonal savings axis no AL framework addresses. AL economizes labels within one model's training; nothing in this literature deduplicates DFT evaluations across models. Lupine's measured 154 shared vs 558 naive evaluations (72.4% fewer on a 29-path panel) stacks multiplicatively with AL's per-model savings: run FLARE/DP-GEN-style uncertainty triggers, but only label the union of cross-model disagreement. The same logic extends small-cell AL (Meng et al. 2025, 100×): Lupine's sparse anchors can be evaluated as small cells around the predicted saddle for another ~2 orders of magnitude, turning the AL "label count" game into a "label count × cell volume" game.
    3. \n
  3. Barrier-focused sparse anchoring vs rare-event MD learning. FLARE's flagship rare-event demonstration (Al vacancy) needed a 1 ns on-the-fly run to learn one migration mechanism; Lupine's 4–6 DFT images near the model-predicted saddle gets the barrier directly. The 2026 foundation-model fine-tuning line (Rensmeyer) independently identifies transition states as the high-value rare events worth oversampling — Lupine can position sparse/union anchoring + theorem-verified saddle geometry as the deterministic, retraining-free answer to the same need.
    4. \n
  4. Replace "label-and-retrain" with "verify-and-patch" to kill AL's documented cost spike. FALCON's honest accounting shows GPR retraining can dominate and even exceed AIMD cost at tight thresholds or high T (full text). Lupine's runtime correction patches uMLIP errors against physical theorems without retraining the base model — directly attacking the cost component (retraining overhead) that AL papers systematically under-report, while keeping a DFT ledger only for theorem-violating regions.
    5. \n
\n

Caveats for the chapter editors. (i) Cross-paper DFT-call comparisons are apples-to-oranges — thresholds, system sizes, and phase-space coverage differ; I flagged context per row. (ii) Two numbers I could not verify are marked [UNVERIFIED] (a-C total database size; Li et al. 2015's factor-30 is verified only as quoted in the Podryabinkin–Shapeev full text). (iii) The MAPbI₃ PRL (arXiv:1903.09613) and stanene npj paper (arXiv:2008.11796) were verified at abstract level only — cited for existence/claims, not for per-system DFT counts. (iv) The Zaverkin (arXiv:2312.01416) "existing methods miss rare events" statement is the field's own most quotable failure admission — useful rhetorically, but it is an abstract-level claim of a 2024 single paper.

\n"}