acceptodds
Under review as a conference paper at ICLR 2027

LiMemFlow: Memory-Enhanced Flow Matching for Lithium-Ion Transport in Solid-State Electrolytes

Abstract

Understanding and predicting long-horizon lithium-ion transport in solid-state electrolytes is important for screening high-conductivity materials. This is especially important near room temperature (300 K), where practical battery operation is most relevant. Although machine-learning interatomic potentials reduce force-evaluation cost, they still rely on sequential small-step integration, motivating large-step generative propagators for long-horizon trajectory prediction. Existing large-step propagators based on flow matching and diffusion models typically formulate trajectory generation as a position-only Markov transition, discarding velocity-induced temporal dependence and collective slow processes that strongly affect diffusion statistics at 300 K. We introduce LiMemFlow, a memory-enhanced flow matching model for lithium-ion trajectory prediction. LiMemFlow combines a PET-based geometric backbone with two complementary memory modules: a point-wise history memory that encodes recent atomic displacement sequences and a pair-wise history memory that stores historical relative motion between neighboring atoms. These memory signals are injected into the equivariant propagator through residual cross-attention and updated during autoregressive rollout. We evaluate LiMemFlow on two trajectory datasets of lithium-containing crystalline structures: our newly computed 300 K dataset and 600–1200 K lithium-transport benchmarks. At 300 K, LiMemFlow reduces MAE by 20.0% and MAE by 17.0% compared with the strongest LiFlow baseline. Consistent with the expected role of memory in slow, path-dependent transport, explicit memory conditioning yields its clearest gains at 300–600 K, including the practically relevant near-room-temperature regime, while the gains diminish at higher temperatures and LiMemFlow remains competitive with the strongest LiFlow baseline at 1000–1200 K.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
flow matching, molecular dynamics, equivariant neural networks, non-Markovian dynamics, history-conditioned models, long-horizon trajectory prediction, lithium-ion transport, solid-state electrolytes

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nnjaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a flow matching model called LiMemFlow which incorporates memory pathway components of point-wise and pair-wise histories. It tackles the problem of long-horizon lithium-ion transport and outperforms previous approaches across several temperature ranges, especially at 300 K. This 300 K dataset is also a contribution of this work.

Critical strengths
  • The results on the long-horizon rollout are encouraging, especially in the newly introduced 300 K temperature range.
  • The ablation of the memory results (Table 2) is thorough, providing a better understanding of the memory components of LiMemFlow.
  • The computational times for LiMemFlow are also very competitive with other models.
Critical weaknesses
  • The writing and presentation style could be improved. As this paper tackles a very specific problem in lithium-ion trajectory prediction, many details will need to be learned by researchers outside this field. This also makes me question the exact suitability of this paper for ICLR. As a note, the previous approaches presented in this paper for comparison are from journals that also allow longer formats.
  • A new benchmark for 300 K temperature is also presented for comparison against previous models, which makes the contributions of the paper in the intersection of new model proposition and new dataset. I am wondering if this paper would also be a better fit for a specific benchmark track at a different venue, in my opinion.
Questions for authors
  • Based on the Appendix, different settings were used for the 300 K and the 600-1200 K models. Can the authors explain the reasoning behind these differences?
  • In Table 2, what does the w/o Memory mean? Does it refer to both the point-wise and pair-wise memory pathways being removed? In the text, it only says "Effect on ranking metrics." on L440, which makes it ambiguous to understand.
  • Can the authors elaborate on the time differences of LiMemFlow compared to LiFlow? Both are very efficient, but it would be helpful to understand which part of the model increases the cost in LiMemFlow.
Reviewer_xG75OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LiMemFlow adds point-wise displacement-history attention and a pair-wise relative-motion memory bank to a PET-based flow-matching propagator (1 ps steps) for Li transport. On a new 974-structure 300 K MLIP-MD dataset it reports a 20.0% lower log-MSD endpoint MAE than LiFlow(P+C), and parity at 1000–1200 K.

Critical strengths

The non-Markovian framing is well motivated. The new 300 K dataset with code is valuable, and the paper includes broad ablations and collective-transport diagnostics (Haven ratio, back-jump ratio). The headline arithmetic is correct.

Critical weaknesses

Weaknesses

  1. No control directly verifies that history is used at inference. The validation-selected inference scale is λ_hist = 10⁻⁵ for 600–1200 K (Table 11), versus training scales of 0.05/0.03 (Table 9). The two scales differ by several orders of magnitude, and no current experiment shows how much the history actually contributes at inference. The paper reports neither the norm of the memory residual relative to the main representation nor results at λ_hist = 0, so it is unclear whether the residual still matters at 10⁻⁵. The 600 K gain over the memory-free model (0.298 vs 0.343) could come from history conditioning or from differences in the trained weights. At 300 K, the validation sweep (Fig. 4a) is sensitive: λ = 10⁻³ gives 0.91 and λ = 0.03 gives 0.65 (validation numbers, not directly comparable with the test-set ablation). Please report, for the same checkpoint, λ_hist = 0, shuffled history, and ground-truth (teacher-forced) history, together with the residual-to-representation norm ratio.

  2. Uncertainty reflects only rollout randomness. Tables 1–2 report ± over three rollout seeds, appendix Tables 3–5 use a single rollout seed, every model is trained once, and there is one reference trajectory per structure. The “statistically indistinguishable” wording at 1000–1200 K should be qualified. Small rollout variance does not rule out training randomness. Given the size of the 300 K gains (20–23%), I do not see this as a reason to reject, but additional training seeds and a reference-vs-reference noise floor would make the results much more convincing. For structures with few hops within 150 ps, the endpoint MSD may mainly reflect local motion; please report the fraction of such structures and the variability between reference trajectories. The MSD setting may need to be read together with a VACF plot to judge whether the chosen time window is reasonable.

  3. It is unclear whether the ablation variants received their own λ_hist sweeps. Given λ_hist's large effect on validation, the ablation comparisons may be biased.

Minor comments

On the baseline reproduction: the official-checkpoint LiFlow(P+C) is slightly weaker than in Table 1 of LiFlow (arXiv:2410.01464v4) at 600–1200 K (0.375/0.396/0.460/0.405 vs 0.348/0.366/0.429/0.389), mainly because the reproduced Corrector gives almost no gain. LiMemFlow still improves ~14% over the published 600 K number and is slightly worse than published at 1000/1200 K. Cross-paper numbers depend on inference settings, so I do not ask for the local parity statement to be withdrawn, but the difference should be noted. Please disclose the reproduction settings and the 300 K baseline tuning in the appendix, and run the baseline several times to remove randomness. Memory slightly lowers Spearman ρ at 300 K and 1200 K, which matters for the screening motivation. MDTrajNet is run without its PES correction; please mark this in the table. Please explain how the NPT reference is reconciled with fixed-lattice generation. Finally, Eq. 4 serves as motivation; the learned modules do not correspond to a memory kernel, so the text should not suggest that they approximate the memory integral.

Questions for authors
  • What are the same-checkpoint inference controls and the residual norm ratio listed in point 1?
  • How was the 300 K baseline tuned?
  • What are the multi-seed results and the reference noise floor?
Reviewer_yByhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LiMemFlow is a flow matching propagator that uses displacement history to predict lithium ion transport in crystalline materials. It combines a PET backbone with pointwise memory of atomic motion and pairwise memory of relative motion between neighboring atoms. Both memories are updated during autoregressive rollout. The paper introduces a MACE dataset of 974 structures with 150 ps trajectories at 300 K. It also evaluates on LiFlow's benchmark at higher temperatures. The appendices examine history length, performance across mobility regimes, and collective transport. At 300 K, the paper reports lower MSD errors than LiFlow(P+C) and better retrieval of high-mobility candidates. The gains diminish at higher temperatures, and the paper attributes this trend to memory mattering most for slow, path-dependent transport.

Critical strengths

The application is well motivated. Transport near room temperature is relevant to electrolyte screening and difficult to sample when hopping events are sparse. The new dataset broadens evaluation beyond LiFlow's benchmark and could support further work on this problem.

The gains at low temperatures are meaningful. At room temperature, LiMemFlow reduces the endpoint log MSD error by about 20% relative to LiFlow(P+C) and improves agreement with the full MSD curve. The paper also reports better retrieval of candidates with high mobility.

There is evidence that history is useful at 300 K. Retrained models with several past displacements perform better than those using the latest displacement alone. In Figure 4a, validation error rises sharply for the same checkpoint as memory strength approaches zero. Component ablations also indicate a larger contribution from pointwise memory within the combined model.

Critical weaknesses

The role of memory at inference is not fully established, and the evidence is weakest at 600 K. In Fig 4b, validation error rises with memory strength and plateaus at the smallest scales. The selected value of 10^-5 is thousands of times below the Table 9 scales that I take to be the training values. This raises doubt about whether memory contributes to the 600 K gain over the memory-free ablation. The gain could instead come from training with the memory modules or from training variance. The figure alone cannot settle this, but the claim of clearest gains at 300-600 K needs evidence that memory is active at 600 K. Table 11 also lists one scale for all higher temperatures, but the text describes separate selections. At 300 K (Fig 4a), the checkpoint clearly depends on memory, but validation error is highly sensitive to the scale on a coarse grid. The gain over the memory-free ablation is much larger on that metric than on the full curve, and rank correlation does not improve. This may reflect calibration of MSD magnitude rather than better modeling of path dependence.

The experiments only partly isolate memory. At 300 K, pairwise memory alone and single-step history are both far worse than no memory. The paper notes that the pathways interact, but it does not explain why pairwise memory alone degrades rollout. The variants use one training seed with no stated retuning of the memory scale, so rollout seed variability understates the uncertainty. Appendix A supports the history settings with test results, and validation selection is documented only for the memory scale. The memory-free model accounts for about half of the curve improvement over LiFlow(P+C), so the headline comparison mixes backbone and memory effects. The comparison between 300 K and the higher temperatures is also confounded by differences in structures, reference potential, rollout horizon, model size, history length, and memory scale. The trend within the higher temperature benchmark is better controlled but depends on the memory scale at each temperature. No analysis measures memory in the reference dynamics directly.

The claim of rotational equivariance is not justified by the architecture as written. The cited PET paper describes PET as not intrinsically equivariant and relies on a separate symmetrization procedure that this paper does not describe. Appendix E projects raw Cartesian components into ordinary hidden features, so the readout coefficients need not be rotation invariant. Expressing the output as a weighted sum of relative vectors is therefore insufficient. The displacement residual raises the same problem. Rotation augmentation provides approximate symmetry at best.

The transport and screening evidence needs clearer interpretation. The largest gains over LiFlow(P+C) occur in mobility tertiles almost entirely below the paper's own mobility threshold. The top tertile gains least, and it nearly coincides with the screening positives. Appendix C still shows errors above an order of magnitude in collective diffusivity and Haven ratios. These diagnostics use a single rollout seed and do not quantify convergence. Against the memory-free ablation, average precision improves, but the recall gain is less than one structure. The stated sensitivity check for the recomputed labels is not reported.

Questions for authors

How does the same checkpoint perform on the test set with memory injection disabled at each temperature? Which residual scale was used for each temperature and ablation? What are the signed log MSD errors as the scale varies, and does the gain persist when history is shuffled in time?

Do the memory gains persist across training seeds and independent reference simulations? Can you report confidence intervals across test structures? Were history length, interval, and layer placement selected on validation data?

What constraints or symmetrization make the velocity predictor equivariant with both memory pathways and the displacement residual included? Can you measure equivariance errors with all inputs rotated consistently?

What preprocessing causes the screening labels to differ between the original index and recomputed trajectories? Do the conclusions hold under both definitions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.