Pith. sign in

REVIEW 2 major objections 6 minor 34 references

A small per-layer adapter lets any equivariant foundation force field respond continuously to charge without rebuilding the model.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 05:48 UTC pith:RKTHPUUN

load-bearing objection Clean, equivariance-preserving FiLM adapter that turns charge conditioning into a few-thousand-frame fine-tune; water results are solid, generality language is a bit ahead of the tests. the 2 major comments →

arxiv 2607.05559 v1 pith:RKTHPUUN submitted 2026-07-06 cs.LG cond-mat.mtrl-sciphysics.chem-phphysics.comp-ph

EquiFiLM: Charge-Conditioned Equivariant Force Fields via Feature-wise Linear Modulation

Samuel Sahel-Schackis , Ken-ichi Nomura , Aiichiro Nakano , Matthias F. Kling , Thomas Linker This is my paper
classification cs.LG cond-mat.mtrl-sciphysics.chem-phphysics.comp-ph
keywords equivariant force fieldsfeature-wise linear modulationcharge conditioningmachine learning interatomic potentialsFiLM adaptermolecular dynamicsliquid water
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Foundation machine-learning force fields map atomic positions and species to energies and forces at near-DFT accuracy, but they treat the electronic state as fixed: two systems that share a geometry yet differ in total charge produce identical predictions. EquiFiLM inserts a lightweight Feature-wise Linear Modulation block at every interaction layer of an equivariant backbone. The block reads a single continuous per-graph scalar (here excess charge density) and uses two small MLPs to scale and shift only the scalar feature channels, leaving higher-rank geometric features untouched so that E(3)-equivariance is preserved exactly. On charged liquid water, the resulting E-MACE model, trained on roughly 6400 DFT frames at four charges, cuts force error by a factor of three and energy error by roughly sixty relative to the same backbone fine-tuned without the adapter, matches a much larger purpose-built charge-aware foundation on forces, generalizes to held-out charges, runs stable molecular dynamics, and recovers the charge-dependent shift of the first oxygen–oxygen shell that ultrafast diffraction would measure. The practical claim is that continuous external control can be added to an existing foundation as an adapter problem rather than a full foundation retraining.

Core claim

EquiFiLM shows that continuous external conditioning of an equivariant foundation force field can be realized by a per-layer FiLM block that modulates only scalar interaction-layer channels from a single per-graph scalar; the construction preserves E(3)-equivariance exactly, adds negligible inference cost, and, when applied to MACE-MatPES on charged liquid water, yields a single set of weights that matches specialist and charge-aware baselines on training charges, generalizes across held-out charges, and supports stable, energy-conserving molecular dynamics whose structural response matches the expected charge-dependent pair-distribution-function shift.

What carries the argument

ChargeFiLMBlock: two small MLPs map the conditioning scalar c to per-channel scale γ(c) and shift β(c) that act only on the scalar (ℓ=0) slice of every message tensor via m′ = (1+γ)⊙m + β, leaving all higher-rank equivariant channels unchanged; zero-initialized so the adapter begins as the identity.

Load-bearing premise

Held-out charge accuracy is measured mainly by re-labeling geometries taken from nearby trained trajectories rather than by fully re-equilibrating nuclei at the new electronic state, so the reported interpolation and extrapolation errors may not capture the true charged potential-energy surface.

What would settle it

Run independent AIMD at a held-out charge (for example q=12e or 18e), recompute forces and energies with the same DFT settings, and check whether E-MACE force RMSE remains inside the claimed 18–61 meV/Å band and whether NVE trajectories stay energy-conserving without special seeding.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The manuscript introduces EquiFiLM, a lightweight per-layer Feature-wise Linear Modulation adapter that injects a continuous per-graph scalar (here total charge, entered as c=q/N) into equivariant foundation MLFFs by scaling and shifting only scalar interaction-layer channels (Eq. 2), thereby preserving E(3)-equivariance exactly. Instantiated on MACE-MatPES for charged liquid water (E-MACE) and trained on ~6400 r2SCAN AIMD frames at four charges, the model reports ~3.1× lower force RMSE and ~61× lower per-atom energy RMSE than an unconditioned fine-tune on the same data, force accuracy competitive with fine-tuned MACE-POLAR-1-M at ~3× lower inference cost, usable held-out interpolation/extrapolation force and energy errors, stable energy-conserving MD, and a charge-dependent first-shell ΔG(r;q) prediction relevant to UED. Ablations (Table 2), a per-state specialist matrix (Appendix A), data-efficiency curves, a GPAW/PBE cross-pipeline force-response check, and a 2×2×2 supercell structural check support the design. Code, data, and a checkpoint are released.

Significance. If the results hold, EquiFiLM is a practically important contribution: it reframes external conditioning of foundation MLFFs as a parameter-efficient adapter problem rather than a from-scratch charge-aware foundation (~10^8 structures) or a bank of per-state specialists. Exact equivariance preservation by scalar-only modulation, zero-init drop-in identity, and negligible inference overhead are clean design points. Strengths that raise confidence include reproducible code and Zenodo data, systematic ablations (β-only, concat embedding, width, init), NVE energy conservation at interpolation and extrapolation charges, falsifiable ΔG(r;q) structural predictions, supercell transfer, and an independent DFT-engine force-difference check. The recipe is of clear interest for electrochemistry, photoinjection, and other driven atomistic processes where foundations currently lack a conditioning axis.

major comments (2)
  1. Section 3.1 and Appendix C (Table 6): held-out force/energy RMSE is measured as vertical electronic response—geometries sampled from AIMD at a nearby training charge and only re-labeled at the target NELECT—not on nuclei re-equilibrated at the new charge. Full charge-plus-geometry generalization is deferred to MD stability and structure. This protocol is stated in the text but is load-bearing for the abstract claim that the model “generalizes” across seven held-out charges and that “one set of weights is usable at any charge.” Please elevate this distinction into the abstract and the opening of §3.1 (e.g., “vertical force response at fixed near-equilibrium nuclei”), and either (i) add a limited set of re-equilibrated single-points or short AIMD-relabeled checks at one interpolation and one extrapolation charge, or (ii) explicitly bound the RMSE claim so readers do not over-read Figure 3
  2. §2.2–§4 and the abstract claim the recipe is backbone- and conditioning-agnostic (“any equivariant MLFF with scalar interaction-layer channels”; temperature, pressure, doping). Empirically only MACE-MatPES + total charge on liquid water is shown. The architectural argument (scalar-only FiLM preserves equivariance) is sound, but the generality claim is stronger than the evidence. Either add a second backbone (e.g., NequIP/Allegro-style or another MACE variant) and/or a second continuous scalar on a small corpus, or rephrase abstract/conclusions to “architecturally applicable; demonstrated on MACE-MatPES with charge,” matching the honest caveats already in §4 about unvalidated γ-dominated axes and multi-axis conditioning.
minor comments (6)
  1. Figure 1 caption and §1: “reaches the accuracy of … MACE-POLAR-1-M fine-tuned” is force-centric; energy comparisons in Figure 3 use offset correction that the text correctly calls generous to POLAR. State “on forces” in the Figure 1 caption to avoid over-reading.
  2. Table 1 vs Appendix B: inference is reported as µs/atom-step with overlapping 1σ bars for E-MACE and fine-tuned MatPES; good. Briefly note in the main text that the pure-functional cat rewrite (Appendix B) is required to avoid IndexPut slowdown, so “indistinguishable cost” assumes the reference implementation.
  3. Appendix C: the jellium-compensated, delocalized excess-electron regime (and the explicit disclaimer against localized hydrated-electron polarons) is important for experimental interpretation of ΔG(r;q). A one-sentence pointer in §3.4.2 would help non-DFT readers.
  4. Notation: c = q/N is introduced in §2.2; some later MD sections switch between q and q_cell without restating that c is held fixed under supercell tiling. A short reminder in Appendix H would reduce confusion.
  5. Related work: SpookyNet and the charge-equilibration lineage are covered; a brief pointer to other conditional/adapter uses of FiLM outside atomistics is already present—fine. Consider citing any concurrent charge-conditioned MACE/UMA adapters if known at revision time.
  6. Typos/clarity: “≈10 8” spacing in abstract/intro; “r 2SCAN” vs “r2SCAN” inconsistency; Figure 3 bottom panel “offset-corrected energy vs. fine-tuned MACE-POLAR-1-M” y-axis label could state units explicitly in the figure.

Circularity Check

0 steps flagged

No significant circularity: EquiFiLM is an empirical adapter whose accuracy claims rest on external DFT labels and independent baselines, not on definitions or self-citation chains that force the reported results.

full rationale

The paper’s load-bearing claims are architectural (scalar-only FiLM preserves E(3)-equivariance by the standard definition of equivariance under O(3)-invariant per-graph gates) and empirical (force/energy RMSE vs no-FiLM, concat embedding, per-state specialists, and fine-tuned MACE-POLAR-1-M; held-out charges; NVE conservation; ΔG(r;q) from MD). Training and evaluation labels are external VASP r2SCAN (with a GPAW/PBE cross-pipeline check). FiLM γ/β are learned parameters, zero-initialized to the identity, not defined to equal the target RMSE or PDF. Ablations and the specialist cross-matrix show the gain comes from per-layer modulation rather than mere access to q. Citations (FiLM, MACE, charge-aware foundations) are external prior art, not author uniqueness theorems that forbid alternatives. The vertical-relabel protocol for held-out charges is a scope limit on what is measured, not a construction that forces the reported errors. No equation or self-citation reduces a claimed prediction to its inputs by definition.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The central claim rests on standard equivariant MLFF machinery plus the design choice that scalar-only FiLM from a per-graph continuous scalar is enough to capture charge-driven PES changes when the backbone already covers the chemistry. Free parameters are ordinary neural training knobs and adapter width; no new physical constants. Invented entities are architectural modules (ChargeFiLMBlock / EquiFiLM), not new physical objects. Domain assumptions about delocalized excess charge and vertical held-out testing bound the scope of the empirical support.

free parameters (4)
  • FiLM MLP hidden width h = h=128 (headline); h=64 (ablations)
    Chosen as 64 for ablations and 128 for headline E-MACE; modest capacity hyperparameter that slightly improves force/energy RMSE at full data.
  • Adapter MLP weights (γ, β networks) = ~0.10 M parameters
    Learned from ~6400 DFT frames; zero-initialized then trained. These are the fitted maps from c to per-channel scale/shift.
  • Training optimizer and schedule (lr, SWA, loss weights) = lr=0.01; SWA 150–250; wF=100
    AdamW lr=0.01, wE=1, wF=100, SWA epochs 150–250; standard but free choices that affect reported RMSE.
  • Conditioning representation c=q/N = c = q/N
    Hand-chosen per-atom charge density so the same c works across cell sizes; not learned from data but defines the input axis.
axioms (5)
  • standard math E(3)-equivariance is preserved if only scalar (ℓ=0) channels are scaled/shifted by an O(3)-invariant function of a per-graph scalar.
    Stated in §2.1–2.2 via the Wigner-D action; standard representation theory of equivariant networks.
  • domain assumption The foundation backbone already covers the chemistry of interest at near-DFT accuracy on the unconditioned baseline.
    Explicit in Discussion: outside that coverage the recipe reduces to ordinary fine-tuning of backbone weights.
  • domain assumption External conditioning (here total charge) acts as a smooth continuous scalar effect on the PES over the studied range.
    Discussion states smoothness in c is assumed; multi-axis and categorical conditioning are untested.
  • domain assumption Jellium-compensated delocalized excess electrons (NELECT = N_neutral + q) are an adequate model of the charge axis for the intended photoinjection/electron-transfer regime.
    Appendix C: setup does not represent localized solvated-electron polarons or explicit counterions.
  • ad hoc to paper Geometries from nearby-charge AIMD re-labeled at target charge adequately probe charge generalization of forces/energies.
    §3.1 and Appendix C protocol; authors note this isolates vertical electronic response, not full re-equilibrated charged ensembles.
invented entities (2)
  • EquiFiLM / ChargeFiLMBlock independent evidence
    purpose: Per-layer adapter that maps a continuous per-graph scalar to γ,β and applies (1+γ)⊙m_scalar + β while leaving ℓ≥1 channels unchanged.
    Architectural invention; independent evidence is empirical (RMSE, MD, PDF) rather than a new physical particle or force. Falsifiable via force/energy benchmarks and dynamics on other systems.
  • E-MACE independent evidence
    purpose: Concrete instance: MACE-MatPES + EquiFiLM trained on charged liquid water.
    Named model artifact; evidence is the reported training/held-out metrics and MD applications.

pith-pipeline@v1.1.0-grok45 · 24802 in / 3851 out tokens · 34132 ms · 2026-07-11T05:48:31.399617+00:00 · methodology

0 comments
read the original abstract

Foundation machine learning force fields (MLFFs) such as MACE-MP-0 and UMA cover broad chemical space at near density functional theory (DFT) accuracy. However, they assume equilibrium ground-state physics and do not natively handle externally induced changes to the electronic state, such as charging, applied fields, or electronic excitation, which limits their use for driven processes such as photoexcitation and charge injection. We propose EquiFiLM, a lightweight extension that adds continuous external conditioning to any equivariant foundation MLFF via a per-layer Feature-wise Linear Modulation (FiLM) block, learning externally driven changes to the potential energy surface from minimal training data. The block modulates only scalar channels and preserves E(3)-equivariance exactly. We demonstrate the recipe on charged liquid water with the foundation model MACE-MatPES as the backbone, yielding E-MACE. On the four training charges, E-MACE delivers a $3.1\times$ reduction in force RMSE ($21.3$ to $6.96$ meV/$\mathring{A}$) and a $61\times$ reduction in per-atom energy RMSE ($6.1$ to $0.1$ meV/atom) over a baseline without EquiFiLM trained on the same data, at indistinguishable inference cost. Across seven held-out interpolation and extrapolation charges, force RMSE stays within $18-61$ meV/$\mathring{A}$ and energy RMSE within $0.7-5.4$ meV/atom. The model runs stable molecular dynamics across the full range tested and predicts the charge-dependent first-shell response of the reduced pair distribution function probed by ultrafast electron diffraction. Adding this conditioning axis to the foundation requires only a few thousand DFT-labeled frames, against the $\approx 10^8$ structures of a charge-aware foundation trained from scratch. The recipe is backbone- and conditioning-agnostic: it applies without architectural change to any equivariant MLFF with scalar interaction-layer channels.

Figures

Figures reproduced from arXiv: 2607.05559 by Aiichiro Nakano, Ken-ichi Nomura, Matthias F. Kling, Samuel Sahel-Schackis, Thomas Linker.

Figure 1
Figure 1. Figure 1: Per-charge force RMSE at the four training charges for seven models across three regimes: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: ChargeFiLMBlock and its integration with MACE. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-charge RMSE against VASP r2SCAN labels; marker shape gives the split (circle: training, square: interpolation, triangle: extrapolation). Top: force RMSE for E-MACE, the no-FiLM baseline (or equivalently fine-tuned MACE-MatPES), and fine-tuned MACE-POLAR-1-M (dashed: chemical accuracy, 1 kcal/mol/Å). Middle: raw per-atom energy vs. the no-FiLM baseline (dashed: thermal energy kBT at 300 K). Bottom: offs… view at source ↗
Figure 4
Figure 4. Figure 4: Training charge mean force RMSE (left axis, blue) and energy RMSE (right axis, red) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Energy conservation over 1.5 ps post-thermostat detach at q = 12e (interpolation, top) and q = 20e (extrapolation, bottom). Epot (red) and Ekin (green) exchange thermal-scale energy of order ±5 meV/atom while their sum Etot (blue) stays within an ≈ 0.18 meV/atom band, with mean drift below 0.01 meV/atom. One set of E-MACE weights gives stable dynamics across the full interpolation-extrapolation range. 3.4.… view at source ↗
Figure 6
Figure 6. Figure 6: Equilibrium charge response of the reduced PDF, [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cross-evaluation heatmaps on the four training charges. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Cross-pipeline per-atom force-difference response on a fixed pool of neutral water geome [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Charge-induced structural response ∆gαβ(r; q) = gαβ(r; q) − gαβ(r; 0) for the three pairs O-O (left), O-H (middle), H-H (right) at q ∈ {4e, 8e, 14e, 18e}. The first O-O coordination shell shifts outward (loss at ≈ 2.7 Å, gain at ≈ 3.3 Å). The O-H and H-H pairs show the same outward redistribution. All features scale monotonically with |q| [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Energy conservation in the 2×2×2 supercell (2, 592 atoms) over a 1.4 ps post-thermostat detach NVE window at q = 12e (interpolation, top) and q = 18e (extrapolation, bottom). As in the 324-atom cell ( [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Per-pair partial pair-correlation differences [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Reduced pair-distribution-function difference [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 6 linked inside Pith

  1. [1]

    A foundation model for atomistic materials chemistry.The Journal of Chemical Physics, 2025

    Ilyes Batatia, Philipp Benner, Yuan Chiang, Gábor Csányi, et al. A foundation model for atomistic materials chemistry.The Journal of Chemical Physics, 2025

    2025

  2. [2]

    Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso- Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, John R

    Brandon M. Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso- Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, John R. Kitchin, Daniel S. Levine, Kyle Michel, Anuroop Sriram, Taco Cohen, Abhishek Das, Ammar Rizvi, Sushree Jagriti Sahoo, Zachary W. Ulissi, and C. Lawrence Zitnick. UMA: A family of universal models for atoms. arXiv prepri...

    arXiv 2025

  3. [3]

    Kaplan, Runze Liu, Ji Qi, Tsz Wai Ko, Bowen Deng, Janosh Riebesell, Gerbrand Ceder, Kristin A

    Aaron D. Kaplan, Runze Liu, Ji Qi, Tsz Wai Ko, Bowen Deng, Janosh Riebesell, Gerbrand Ceder, Kristin A. Persson, and Shyue Ping Ong. A foundational potential energy surface dataset for materials.arXiv preprint arXiv:2503.04070, 2025

    arXiv 2025

  4. [4]

    Hoffman, Sabrina C

    Yi-Lun Liao, Alexander J. Hoffman, Sabrina C. Shen, Alexandre Duval, Sam Walton Norwood, and Tess Smidt. EquiformerV3: Scaling efficient, expressive, and general SE(3)-equivariant graph attention transformers.arXiv preprint arXiv:2604.09130, 2026

    Pith/arXiv arXiv 2026

  5. [5]

    Wood, Luis Barroso-Luque, Daniel S

    Xiang Fu, Brandon M. Wood, Luis Barroso-Luque, Daniel S. Levine, Meng Gao, Misko Dzamba, and C. Lawrence Zitnick. Learning smooth and expressive interatomic potentials for physical property prediction.arXiv preprint arXiv:2502.12147, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Lee, and Seungwu Han

    Jaesun Kim, Jinmu You, Yutack Park, Yunsung Lim, Yujin Kang, Jisu Kim, Haekwan Jeon, Suyeon Ju, Deokgi Hong, Seung Yul Lee, Saerom Choi, Yongdeok Kim, Jae W. Lee, and Seungwu Han. Optimizing cross-domain transfer for universal machine learning interatomic potentials.arXiv preprint arXiv:2510.11241, 2025

    arXiv 2025

  7. [7]

    Orb-v3: atomistic simulation at scale.arXiv preprint arXiv:2504.06231, 2025

    Benjamin Rhodes, Sander Vandenhaute, Vaidotas Šimkus, James Gin, Jonathan Godwin, Tim Duignan, and Mark Neumann. Orb-v3: atomistic simulation at scale.arXiv preprint arXiv:2504.06231, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    PET-MAD as a lightweight universal interatomic potential for advanced materials modeling.Nature Communications, 16:10653, 2025

    Arslan Mazitov, Filippo Bigi, Matthias Kellner, Paolo Pegolo, Davide Tisi, Guillaume Fraux, Sergey Pozdnyakov, Philip Loche, and Michele Ceriotti. PET-MAD as a lightweight universal interatomic potential for advanced materials modeling.Nature Communications, 16:10653, 2025

    2025

  9. [9]

    Baldwin, Domantas Kuryla, Joseph Hart, Elliott Kasoar, Alin M

    Ilyes Batatia, William J. Baldwin, Domantas Kuryla, Joseph Hart, Elliott Kasoar, Alin M. Elena, Harry Moore, Mikołaj J. Gawkowski, Benjamin X. Shi, Venkat Kapil, Panagiotis Kourtis, Ioan- Bogdan Magd˘au, and Gábor Csányi. MACE-POLAR-1: A polarisable electrostatic foundation model for molecular chemistry.arXiv preprint arXiv:2602.19411, 2026

    arXiv 2026

  10. [10]

    Finkler, Stefan Goedecker, and Jörg Behler

    Tsz Wai Ko, Jonas A. Finkler, Stefan Goedecker, and Jörg Behler. A fourth-generation high- dimensional neural network potential with accurate electrostatics including non-local charge transfer.Nature Communications, 12:398, 2021

    2021

  11. [11]

    Anstine, Maike Bergeler, V olker Settels, Con- rad Stork, Sebastian Spicher, and Olexandr Isayev

    Bhupalee Kalita, Roman Zubatyuk, Dylan M. Anstine, Maike Bergeler, V olker Settels, Con- rad Stork, Sebastian Spicher, and Olexandr Isayev. AIMNet2-NSE: A transferable reac- tive neural network potential for open-shell chemistry.ChemRxiv, 2025. doi: 10.26434/ chemrxiv-2025-kdg6n-v2

    2025

  12. [12]

    Panagiotopoulos, Roberto Car, and Weinan E

    Linfeng Zhang, Han Wang, Maria Carolina Muniz, Athanassios Z. Panagiotopoulos, Roberto Car, and Weinan E. A deep potential model with long-range electrostatic interactions.Journal of Chemical Physics, 156:124107, 2022

    2022

  13. [13]

    Unke, Stefan Chmiela, Michael Gastegger, Kristof T

    Oliver T. Unke, Stefan Chmiela, Michael Gastegger, Kristof T. Schütt, Huziel E. Sauceda, and Klaus-Robert Müller. SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects.Nature Communications, 12:7273, 2021

    2021

  14. [14]

    FiLM: Visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. FiLM: Visual reasoning with a general conditioning layer. InAAAI Conference on Artificial Intelligence, 2018. 11

    2018

  15. [15]

    Ilyes Batatia, Dávid Péter Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. InAdvances in Neural Information Processing Systems, 2022

    2022

  16. [16]

    Atomic cluster expansion for accurate and transferable interatomic potentials

    Ralf Drautz. Atomic cluster expansion for accurate and transferable interatomic potentials. Physical Review B, 99:014104, 2019

    2019

  17. [17]

    Kresse and J

    G. Kresse and J. Furthmüller. Efficient iterative schemes forab initiototal-energy calculations using a plane-wave basis set.Physical Review B, 54:11169–11186, 1996

    1996

  18. [18]

    Furness, Aaron D

    James W. Furness, Aaron D. Kaplan, Jinliang Ning, John P. Perdew, and Jianwei Sun. Accurate and numerically efficient r2SCAN meta-generalized gradient approximation.Journal of Physical Chemistry Letters, 11:8208–8215, 2020

    2020

  19. [19]

    Averaging weights leads to wider optima and better generalization.arXiv preprint arXiv:1803.05407, 2018

    Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization.arXiv preprint arXiv:1803.05407, 2018

    Pith/arXiv arXiv 2018

  20. [20]

    Ming-Fu Lin, Nirmal Singh, Senmiao Liang, Mianzhen Mo, J. P. F. Nunes, Kathryn Ledbetter, Jie Yang, Michael Kozina, Stephen Weathersby, Xiaozhe Shen, et al. Imaging the short-lived hydroxyl-hydronium pair in ionized liquid water.Science, 374:92–95, 2021

    2021

  21. [21]

    E(3)-equivariant graph neural networks for data-efficient and accurate inter- atomic potentials.Nature Communications, 13:2453, 2022

    Simon Batzner. E(3)-equivariant graph neural networks for data-efficient and accurate inter- atomic potentials.Nature Communications, 13:2453, 2022

    2022

  22. [22]

    Owen, Mordechai Kornbluth, and Boris Kozinsky

    Albert Musaelian, Simon Batzner, Anders Johansson, Lixin Sun, Cameron J. Owen, Mordechai Kornbluth, and Boris Kozinsky. Learning local equivariant representations for large-scale atomistic dynamics.Nature Communications, 14:579, 2023

    2023

  23. [23]

    Courville

    Harm de Vries, Florian Strub, Jérémie Mary, Hugo Larochelle, Olivier Pietquin, and Aaron C. Courville. Modulating early visual processing by language. InAdvances in Neural Information Processing Systems, 2017

    2017

  24. [24]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022

    2022

  25. [25]

    Parameter-efficient transfer learning for NLP

    Neil Houlsby, Andrei Giurgiu, Stanisław Jastrz˛ ebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. InInternational Conference on Machine Learning, 2019

    2019

  26. [26]

    David Ha, Andrew Dai, and Quoc V . Le. HyperNetworks. InInternational Conference on Learning Representations, 2017

    2017

  27. [27]

    Smith, Benjamin T

    Roman Zubatyuk, Justin S. Smith, Benjamin T. Nebgen, Sergei Tretiak, and Olexandr Isayev. Teaching a neural network to attach and detach electrons from molecules.Nature Communica- tions, 12:4870, 2021

    2021

  28. [28]

    Probing the self-ionization of liquid water with ab initio deep potential molecular dynamics.Proceedings of the National Academy of Sciences, 120:e2302468120, 2023

    Marcos Calegari Andrade, Roberto Car, and Annabella Selloni. Probing the self-ionization of liquid water with ab initio deep potential molecular dynamics.Proceedings of the National Academy of Sciences, 120:e2302468120, 2023

    2023

  29. [29]

    Enhanced deep potential model for fast and accurate molecular dynamics: application to the hydrated electron.Physical Chemistry Chemical Physics, 26:23080–23088, 2024

    Ruiqi Gao, Yifan Li, and Roberto Car. Enhanced deep potential model for fast and accurate molecular dynamics: application to the hydrated electron.Physical Chemistry Chemical Physics, 26:23080–23088, 2024

    2024

  30. [30]

    A foundation machine learning potential with polarizable long-range interactions for materials modelling.Nature Communications, 16:10484, 2025

    Rongzhi Gao, ChiYung Yam, Jianjun Mao, Shuguang Chen, GuanHua Chen, and Ziyang Hu. A foundation machine learning potential with polarizable long-range interactions for materials modelling.Nature Communications, 16:10484, 2025

    2025

  31. [31]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  32. [32]

    Reddi, Satyen Kale, and Sanjiv Kumar

    Sashank J. Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of Adam and beyond. arXiv preprint arXiv:1904.09237, 2019. 12

    Pith/arXiv arXiv 1904

  33. [33]

    excess electron

    Jens Enkovaara, C. Rostgaard, Jens J. Mortensen, J. Chen, M. Dułak, L. Ferrighi, J. Gavnholt, C. Glinsvad, V . Haikola, H. A. Hansen, et al. Electronic structure calculations with GPAW: a real-space implementation of the projector augmented-wave method.Journal of Physics: Condensed Matter, 22:253202, 2010. 13 Appendix A Per-state cross-evaluation matrix T...

    2010

  34. [34]

    Odd-q configurations would require open-shell treatment and are out of scope. In a periodic cell the added electrons are compensated by a uniform neutralizing background (the standard convention for charged supercells), so the excess charge is delocalized across the cell rather than localized as a solvated electron with an explicit counter-cation. The con...