acceptodds
Under review as a conference paper at ICLR 2027

SPLICE: Sparse Physics Library Identification and Compositional Expansion

Abstract

Neural operator splitting composes a library of neural operators, one per physical mechanism, into zero-shot simulators of new equations. Existing approaches, however, require these mechanisms and their labels to be specified in advance. We ask whether a pretrained library can instead be adapted and grown from unlabeled trajectories in which multiple mechanisms act simultaneously. We introduce SPLICE, a framework that discovers operators without mechanism labels on mixture data. The contributions of the candidate operators to each trajectory are inferred by solving a convex optimization problem. A calibrated test then assigns each discovered operator the mechanism it implements. Failure cases are also detectable, when the library lacks the required physics or when mechanisms are insufficiently separable. Across one and two dimensional systems, the discovered operators recover mechanisms from unlabeled mixtures and compose into zero-shot simulators of unseen equations; on Navier-Stokes they compose about as accurately as their supervised counterparts on average.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, operator splitting, partial differential equations, zero-shot, self-supervised learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_cWsUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces SPLICE, a method for adapting a pretrained operator library to unlabeled mixture trajectories. Each library entry is a learned time-stepper: an encoder maps a context window to a latent code, a hypernetwork maps that code to vector-field weights, and RK4 advances one step. Discovery alternates an exact convex assignment step with a gradient refinement step, and extends the library by reseeding when residuals persist. Each discovered operator is then graded post hoc against reference solvers using a displacement distance on a fixed probe set. Experiments cover 1D mixtures, coupled Navier-Stokes, and transfer from The Well.

Critical strengths

A key strength is that identification is falsifiable and separate from training. Discovered operators are compared with reference solvers on a fixed probe set, with a margin over the second-closest mechanism and a null calibrated from wrong-kind operators. Because labels are withheld, this tests real identification, not renamed decomposition. That is a genuine conceptual contribution.

The strongest evidence is coupled Navier-Stokes, where no supervised data exist. Euler and diffusion act together at every step, yet the discovered library identifies both on all 10 seeds and composes at 0.0116 plus or minus 0.0014, about 32 percent better than the 0.017 persistence baseline. In this setting, label-free adaptation stands in for supervised training rather than competing with it. Pure mechanisms show it can match supervised training when both are possible; coupled flow shows a result that supervised training cannot provide, because no labeled data exist.

Critical weaknesses

The abstract says the method adds operators for dynamics the library cannot express, but the paper itself says adaptation cannot introduce a mechanism absent from the initial library. Reseeding can only recover mechanisms the pretrained network already represents, not mechanisms absent from it, so the grow claim conflates not active with not represented.

The coupled-flow result lacks a composition-only baseline. Coupled flows beat persistence, but the paper reports no supervised row and no baseline that composes pretrained mechanisms without discovery. Readers cannot tell how much of the gain comes from composition and how much from discovery. Pure mechanisms have a supervised row; coupled flows do not.

Questions for authors

Given Appendix C.2, should the “adds an operator” claim be restricted to mechanisms already represented but inactive?

On coupled flows, can you report a baseline that composes the pretrained mechanisms without any discovery, so readers can see how much of the gain comes from composition rather than discovery?

Each 2D discovery seed is graded on its own probe set and null. How is cross-seed identification assessed?

1 comment

Thank you for recognizing the value of learning reusable operators from trajectories where several mechanisms always act together.

Identity is not constrained during learning. This is a fair point. A lower prediction error does not by itself mean the mechanisms are separated, which is why we grade identity against reference solvers rather than infer it from accuracy. We also tested using that grading as a constraint on coupled Navier–Stokes. Starting from the library discovered on pure mixtures, we run the adaptation rounds and check each operator against the reference solvers after every round. If an update makes an operator lose its identity, the update is undone. Identification then constrains learning rather than only evaluating it afterwards. Over three seeds, this keeps the advective operator from drifting toward diffusion, at no cost in accuracy (0.157 to 0.184 against 0.183 without it, on the 12 held-out viscosities; Appendix C.3, “Identity gate”). For discovery from coupled flows, we also graded every adaptation round. Both mechanisms are identified on all ten seeds at each of the eight rounds, with positive margins throughout (Appendix C.3, “Discovery at native 256²”).

When Grow can recover a missing mechanism (Appendix C.2)., and when it works. Reseeding creates a new code for the same frozen network, so it can only recover a mechanism that network can already produce. When the network was pretrained with Euler, reseeding regrows a missing Euler operator on three of three seeds (§5.4). On the network pretrained without Euler, it cannot. Even a single code fitted with labels on pure Euler windows still identifies as diffusion (Appendix C.2). A mechanism the network cannot produce would therefore require new parameters rather than a new code. The revised text states both conditions, and claims growth only for mechanisms the pretrained network can already produce.


Reviewer cWsU

We thank the reviewer for the careful review, and for seeing the falsifiable identification protocol as a conceptual contribution.

Adding mechanisms that are not represented. Yes, and we revised the abstract and §3.3 accordingly. Reseeding recovers mechanisms that the pretrained network can produce but that are inactive in the library. It cannot add a mechanism the network does not represent, as Appendix C.2 shows. The revised claim is restricted to the first case, and we state the second as a limitation, since it would require a new operator with its own parameters.

Composition-only baseline on coupled flows. The supervised library is this baseline, since it composes operators pretrained on pure mechanisms without any discovery. Following your request, we extended the Navier–Stokes evaluation to the coupled-flow test set over 12 held-out viscosities and 16 steps. This also corrects the test split used in the submitted Tables 3 and 12, which changes the scale of every number, including the 0.0116 you cite. On the corrected split the coupled-flow result is stronger, 0.139 against 0.741 for persistence. Discovery from coupled flows starts from the supervised library, so the gain from 0.159 to 0.139 is what discovery adds on top of composition. We also add NOS trained with one operator per viscosity as a supervised upper bound. It is trained with a label for each viscosity seen at test time, whereas our supervised library uses mechanism labels only and our label-free libraries use no labels during discovery. Appendix D defines each row.

PersistenceNOS, one operator per viscositySupervisedLabel-free, pure mixturesLabel-free, coupled flows
NRMSE0.7410.0590.159 ± 0.0150.177 ± 0.0100.139 ± 0.017
Identified10/1010/10

Mean ± std over 10 seeds (3 for supervised). NOS and persistence are single models.

Cross-seed identification. For the Navier–Stokes results of Table 3, all seeds are now graded on one shared probe set and one shared null. Both mechanisms are identified on all ten seeds, for both label-free libraries, and this holds when we redraw the null five times (Appendix B.1).

Reviewer_oJsUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the SPLICE framework, which can adapt and expand the neural operator library from unlabeled hybrid physical trajectories, with the aim of reducing the reliance of existing neural operator splitting methods on predefined physical mechanisms and single-mechanism labeled data. SPLICE attempts to recover neural operators with independent physical significance and the ability to be reused from the observed trajectories where multiple physical mechanisms act simultaneously. The author conducted experiments on one-dimensional and two-dimensional physical systems, including coupling Navier-Stokes dynamics. The experimental results show that SPLICE can restore hidden physical mechanisms in multiple Settings and recombine the discovered operators to simulate physical equations that have not been seen before during the training process. In some experiments, the operator library obtained through unlabeled discovery can achieve simulation accuracy similar to that of the supervised training operator library.

Critical strengths
  1. This paper takes into account that in actual physical systems, multiple physical mechanisms often act simultaneously, making it difficult to obtain observational data when each mechanism operates independently. To address this issue, the author proposed the SPLICE framework, which discovers and learns reusable physical operators from unlabeled mixed trajectories through operator adaptation, residual-driven operator library expansion, and independent physical mechanism identification. Compared with the existing neural operator splitting methods that rely on predefined physical mechanisms and single-mechanism labeled data, this paper attempts to construct an operator library that can be continuously expanded and combined, which has clear research value.
  2. The author not only verified the mechanism's recovery ability through one-dimensional and two-dimensional physical systems, but also evaluated the physical identities of the discovered operators using independent reference solvers, and investigated the combined predictive ability of these operators on the coupled Navier-Stokes system. In addition, the paper provides supplementary experiments on different mechanism recognition strategies, operator library expansion methods, and network architectures, and analyzes the reproduction and comparison Settings of related methods. Especially in some Navier-Stokes experiments, the discovered operator library can achieve a prediction accuracy similar to that of the supervised operator library, providing experimental support for the feasibility of learning composable physical operators from unlabeled mixed trajectories.
Critical weaknesses

SPLICE optimizes the encoding of the discovery operator to adapt to unlabeled mixed physical trajectories. However, during the optimization process, the trajectory prediction error is mainly relied on. Therefore, a better prediction result does not necessarily mean that the model has successfully separated different physical mechanisms. The experiments in Appendix C.3 also indicate that in some cases, the discovered Euler operator gradually leans towards the effect of diffusion, leading to the failure of mechanism identity recognition. Although the author verified the physical identity of the discovery operator using an independent reference solver and discussed the related limitations in the paper, this recognition process was mainly used for post hoc evaluation and did not directly constrain the learning process of the operator. Therefore, the existing methods cannot fully guarantee that the discovery operator maintains an independent and stable physical identity during the label-free adaptation process. This, to a certain extent, has affected the persuasiveness of the paper regarding the discovery of physical mechanisms and the ability to reuse operators.

Questions for authors

Appendix C.2 indicates that when the initial operator library lacks the Euler mechanism, the model can restore diffusion but fails to further discover Euler. May I ask what the main cause of this phenomenon is? Under what conditions can the Grow module effectively discover missing physical mechanisms?

1 comment

We thank the reviewer for recognizing the value of learning reusable operators from trajectories containing several acting mechanisms.

Identity is not constrained during learning. This is a fair point. A lower prediction error does not by itself mean the mechanisms are separated, which is why we grade identity against reference solvers rather than infer it from accuracy. We also tested using that grading as a constraint on coupled Navier–Stokes. Starting from the library discovered on pure mixtures, we run the adaptation rounds and check each operator against the reference solvers after every round. If an update makes an operator lose its identity, the update is undone. Identification then constrains learning rather than only evaluating it afterwards. Over three seeds, this keeps the advective operator from drifting toward diffusion, at no cost in accuracy ((0.157 to 0.184 against 0.183 without it, on the revised Navier-Stokes evaluation of Table 3; Appendix C.3, “Identity gate”). For discovery from coupled flows, we also graded every adaptation round. Both mechanisms are identified on all ten seeds at each of the eight rounds, with positive margins throughout (Appendix C.3, “Discovery at native 256²”).

When Grow can recover a missing mechanism (Appendix C.2). Reseeding creates a new code for the same frozen network, so it can only recover a mechanism that network can already produce. When the network was pretrained with Euler, reseeding regrows a missing Euler operator on three of three seeds (§5.4). On the network pretrained without Euler, it cannot. Even a single code fitted with labels on pure Euler windows still identifies as diffusion (Appendix C.2). A mechanism the network cannot produce would therefore require new parameters rather than a new code. The revised text states both conditions, and claims growth only for mechanisms the pretrained network can already produce.

Reviewer_YGrJOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a framework for learning and composing neural operators associated with different physical mechanisms in PDEs. Building upon Neural Operator Splitting (NOS), the method employs Strang splitting to combine representing individual mechanisms. The main contribution is a strategy to discover and adapt these operators from trakectories in which the underlying physical mechanisms are not explicitly labeled.

The pipeline is decomposed in four stages: Build, Adapt, Grow and Identify. In the Build stage, an initial library of neural operators is trained by using synthetic data generated from individual known mechanisms (such as advection and diffusion). The Adapt stage aims to discover the mechanism involved in mixed trajectories. Interestingly, this is accomplished by alternating between a convex optimization problem, which determines the contribution of each operator, and a non-convex optimization step, which adjusts their latent representations. Grow stage allows the library to be expanded when the existing operators cannot adequately represent the observed dynamics. Finally, in the Identify stage, the physical meaning of each learned operator is inferred by comparing its behavior with reference physical solvers.

Critical strengths

The possibility of identifying mechanisms acting simutaneously is a meaningful extension of NOS. The Navier-Stokes experiments are concrete evidence that this strategy may work.

The strategy of alternating a convex problem to estimate the contribution of the operators with a non-convex problem to adjust their latent representations is also interesting (the ablation studies show that this is better than amortized assignment). This is a solid methodological contribution.

The experiments are carefully documented, and the limitations are openly discussed.

Critical weaknesses

In my view, there are two relevant weaknesses, both discussed by the authors: first, the fact that the mechanisms must be significantly distinct for achieving correct identification. Second is that the correct identification of physical mechanisms does not necessarily give precise compositions. Both limit the applicability of the method.

Questions for authors

As far as I understood, you work solely with PDEs where only first-order time derivatives appear. Have you considered extending to second-order evolution problems, such as the wave equation? Sure, you can formulate them as first order systems, but do you foresee any difficulties in this scenario?

3 comments

We thank the reviewer for the positive assessment and refinement steps.

Mechanisms must be distinct. We agree this is a real condition, and the method is built to detect when it fails. Identification requires a margin over the second-closest mechanism and over a null calibrated from wrong-kind operators. Appendix C.3 shows this in practice. When the discovered Euler operator drifts toward diffusion, it loses its Euler label instead of keeping it. §5.5 states this condition as the first limit of the method.

Identification does not guarantee precise composition. We agree, and the paper treats them as separate properties. Identification tells what each operator learned, measured against reference solvers, while composition accuracy is measured by rollout error, and we report both. §5.5 discusses the cases where they differ.

Second-order equations. Writing the wave equation as a first-order system in (u,∂tu)(u, \partial_t u) fits the framework, with operators acting on the pair. We expect the main difficulty to be splitting error on oscillatory, energy-conserving dynamics, where Strang splitting accumulates phase error and a symplectic splitting would be the natural choice. Identification needs only a reference solver for each candidate operator, which is straightforward here. We added this as a direction in the conclusion.

Thank you very much for the explanations! Your paper deals with a very exciting topic in SciML, it presents a clear advance, and the limitations are well documented and discussed. For that reasons, I'm updating my rating.

Thank you for your follow-up and for raising your score. We are glad the explanations were helpful, and we appreciate the time you took to engage with the paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General comment

We thank the reviewers for their careful reading and constructive questions. The revised PDF makes four changes, each answering a point raised below.

  • The Grow claim is restricted to mechanisms the pretrained network represents (abstract, §3.3, §5.5, cWsU and oJsU).
  • Identification across seeds uses one shared probe set and null (Appendix B.1, cWsU).
  • A new experiment uses identification to constrain adaptation (Appendix C.3, oJsU).
  • The Navier–Stokes evaluation is revised and adds a supervised upper bound (Table 3, Appendix D, cWsU).

Dear authors and reviewers,

The author-reviewer discussion period closes on November 18 (AoE). Please engage/continue to engage during this period so that we have our issues discussed thoroughly.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.