WorksComputer Science
Exponential-Nemytskii Neural Operators: Learning Field Evolution Across Step Sizes
Abstract
In variable-step operator learning, agreement at observed step sizes does not by itself determine predictions at other temporal resolutions. We propose an operator whose step-size dependence is prescribed by a first-order Duhamel approximation of semilinear dynamics. It combines exact linear propagation with a local nonlinear function of the field and its derivatives, allowing evaluation below the observation interval without supervision densely sampled across step sizes. Under suitable conditions and at fixed spatial resolution, we prove that agreement between models at two step sizes controls their prediction differences throughout a prescribed range. We also give conditions for local identifiability of both components from two step sizes, supporting interpretation of the learned structure. A latent extension applies the same construction to learned fields. We evaluate the base model on two synthetic problems and the extension on two real datasets against several baselines.
Introduction
Many quantities of scientific and practical interest are continuous fields that evolve in space and time: the elevation of a water surface, the concentration of a diffusing species, the liquid-water content of a storm seen by radar [Veillette et al., 2020], or the radiative power of a spreading wildfire [Wooster et al., 2005]. The continuous field itself is rarely observed directly. Instead, we typically rely on a finite set of noisy samples taken on a spatial grid at discrete observation times. The task is to learn the underlying spatiotemporal evolution from these samples and to predict the field’s future states. Operator learning [8] (Lu et al. 2021) has emerged as a popular approach to this task. Rather than fitting a discrete mapping between fixed-size arrays, it learns the solution operator that advances the field in time as a mapping between function spaces. Because the learned operator acts on the underlying functions rather than their discrete samples, it can be evaluated on spatial grids other than those used during training, and it can be applied to complex systems even when their governing equations remain unknown. The other family of methods are equation learning [1], which focus on recovering governing equations from the data. Our work focuses on operator learning; equation learning is discussed further in § 2.4.
While operator learning treats space as continuous, many autoregressive models, including FNO-2D [Li et al., 2021], the autoregressive version of DeepONet [Lu et al., 2022] and KNO [Xiong et al., 2024], advance the field at a fixed time interval in both training and prediction. The step size is thereby built into the learned map, which does not directly support prediction at arbitrary temporal intervals. Considerable effort has been devoted to developing methods for variable-step prediction, and we review these approaches in § 2.4. In this work, we propose a new operator-learning method for variable-step prediction, based on a first-order Duhamel approximation of semilinear dynamics. We combine this construction with multi-step-size training and establish conditions under which prediction maps at two distinct step sizes locally identify the linear and nonlinear components. Our main contributions are as follows.
Variable-step prediction with a stability guarantee. Our model’s dependence on the step size is prescribed by its mathematical construction, rather than learned entirely from data. This structure allows us to establish a theoretical stability guarantee for prediction across temporal resolutions, without requiring training data densely sampled across step sizes. We prove that, under stated conditions, close agreement with a reference model at just two distinct training step sizes controls prediction differences throughout a prescribed range of step sizes, including below the observation interval (Theorem 1).
Identifiability and interpretability. When applied directly to the observed fields, our model explicitly learns a linear differential operator and a local nonlinear function of the field and its derivatives. These components can be inspected directly to understand how linear propagation and nonlinear effects contribute to the prediction. This interpretation is ambiguous when different choices of components produce identical predictions. We prove that, under suitable conditions, the linear and nonlinear components of our model are locally identifiable when trained at two distinct step sizes (Theorem 3).
We test the method on two synthetic problems and on two real datasets, and compare it against a variety of baselines from operator learning and equation learning.
Method
Setup and model derivation
We observe a spatio–temporal field , , , only through discrete, possibly noisy samples
taken at spatial nodes and times . Our goal is to learn the underlying model of the field and use it to make predictions.
We assume that is described by a semilinear evolution equation,
where is an unknown linear operator generating a strongly continuous semigroup on a Banach space , and is an unknown, locally Lipschitz nonlinear operator. Our aim is to learn both from Eq. (1) alone, in the form of a neural operator that predicts at later times.
We construct our learning model using the Duhamel representation of Eq. (2), which does not merely split the dynamics into a linear part and a correction, but prescribes how the two combine. Write
for the purely linear evolution of the initial state. A mild solution of Eq. (2) satisfies the Duhamel formula
and is the fixed point of , where , constructed by Picard iteration from . Successive Picard iterates differ by , which gains one order in at each step since is Lipschitz, so only the first correction enters at first order. Keeping it alone, , and expanding near (substitute , then use and ) gives, at ,
Over a short horizon the solution is the exactly propagated linear flow plus a single local correction evaluated at the current linear state. Eq. (5) is the model we train to predict forward in time.
Parametrisation and training
To use Eq. (5) we must learn and from data. We parametrise each by a structured class matching its role, rather than by a generic neural network.
Spectral linear propagator. We assume is a constant-coefficient differential operator of order at most on a periodic domain, with and . The Fourier transform replaces by , so is a Fourier multiplier with symbol
a polynomial in whose coefficients are those of itself; in one dimension . The linear flow Eq. (3) becomes with , and the coefficients are the trainable parameters. The propagator is the exact solution of the linear part, so the truncation in Eq. (5) is attributable to the nonlinear term alone. In the implementation we truncate the frequency modes to , where is a hyperparameter (values in Appendix § 7.§ 2.1).
For a bound uniform in , we enforce for all , so that : the linear flow is bounded on at every step size and every discretisation. This is imposed by the following construction. Only terms of even total order contribute to ; the odd-order terms, which carry transport and dispersion, are purely imaginary and are left free, as is the constant term . Among the even-order terms, those in which every appears to an even power (, , and so on) are nonnegative everywhere and receive a nonpositive coefficient, so they can only push down; the rest, such as , change sign with direction and are dropped.
Nemytskii operator. For the nonlinear part of Eq. (2) we do not consider nonlocal nonlinearities, and assume instead that is local of some order : the value of at depends on only through and its spatial derivatives at up to order (the -jet of at ). Write
for the feature map, which collects the field, its spatial derivatives up to order , and, optionally, externally given fields ; its number of entries is . We then assume
so that is a generalised Nemytskii (superposition) operator [9] induced by the finite-dimensional map . Accordingly, the goal becomes to learn the finite-dimensional function and we represent it by a multilayer perceptron (MLP) with smooth activations. The derivatives in are evaluated spectrally, .
Exponential-Nemytskii neural operator. Combining the two components using Eq. (5),
with trainable objects and . We call the resulting model the Exponential–Nemytskii Neural Operator (ENNO), reflecting its combination of exponential linear propagation and a learned Nemytskii map acting on local field features. Predictions over horizons longer than a single first-order step are obtained by applying Eq. (9) autoregressively: starting from ,
is the prediction at time .
Training. In training we use only the one-step model Eq. (9); the multi-step composition Eq. (10) is used only for prediction. A key feature of our approach is that the model is trained with several step sizes , which is required for the identifiability of (§ 2.3.2). Concretely, fix step sizes , , chosen among the differences of observation times, and for each let be the set of observation times for which is also an observation time. For each , we apply the one step of size and compare with the observation at ; averaging the squared error over the grid and over gives
and the training objective is their sum, . Since enters explicitly, the same is used for every . With a single step size the constant term of and the linear part of cannot be separated (Lemma 2); a second step size removes this ambiguity (Theorem 3(ii)).
Latent fields. An extension of the method handles the case where does not follow the semilinear model Eq. (2): we map into a latent field, or several latent fields, using an encoding–decoding structure. Namely, we use an encoder to map the observed history into the latent fields , advance all of them with the same and , and use a decoder to map the latent state back:
Here we restrict both and to be Nemytskii operators, generated by pointwise maps and , so every learnable object in the model is generated by a map between finite-dimensional spaces. The encoder and decoder are trained jointly with and . Taking acts as a delay embedding: when is a partial observation of a higher-dimensional state, a single frame need not determine the subsequent evolution. Only the initial latent state uses the history; the rollout remains Markovian in . Note that and are shared across the latent fields, so the parameter count does not grow with : the fields evolve independently under the same dynamics but from different initial channels produced by , and are recombined nonlinearly by . The price is interpretability: in the base model and describe the mechanisms governing itself, whereas here they describe the latent fields and no longer correspond directly to the equation for .
Theory
Stability across step sizes
We work at fixed spatial resolution on a periodic box . Let be the real Fourier space with integer mode indices , let be its orthogonal projection, and write . All operator norms below are induced by the norm. For , let have symbol as in Eq. (6), and set and . Here is a spatially homogeneous, locally Lipschitz scalar function, and collects the field and its distinct partial derivatives through order . For a second pair from the same class, define , , and analogously. Let denote the real-preserving Fourier multipliers on , and define the nonlinear contrast . We have the following theorem where the proof is given in Appendix § 7.§ 1.
Theorem 1 (Cross-step stability). Fix , and suppose . Assume that there exist and such that
Let be a Lipschitz constant for on . There exist , , and , depending only on , such that for any , setting
we have, for every ,
The linear dependence on this ratio is sharp in general.
Interpretation. The theorem can be understood as follows. If a model’s predictions are close to those of a reference model at two training step sizes, their prediction difference per unit time also has an explicit upper bound at other step sizes. At intermediate step sizes, the same theoretical bound applies as at the training steps. Outside this interval, the bound grows linearly with the extrapolation distance relative to the separation between the training steps. This provides a stability guarantee for using the same model at different temporal resolutions.
Identifiability
We study structural identifiability of from the one-step prediction maps. Write , where lists each distinct partial derivative through order once, including the field value. Accordingly, write , with the field value. Let be continuous and . Proofs of the two results below are in Appendix A.1.2.
Lemma 2 (Single-step degeneracy). Fix . For , set and , and define
Then
whenever the maps are defined.
Theorem 3 (Identifiability). Let and have symbols in the real-coefficient polynomial class of Eq. (6) and nonlinearities that are in and continuous in . Fix and , with the sufficient threshold specified in Appendix A.1.2. At every step size considered, assume for and all with .
(i) One step size. Let . Suppose on for every with , and for some . Then, for some and , and on with , where is defined by Eq. (15). In particular, for .*
(ii) Two step sizes. Under the hypotheses of (i), suppose also on for the same , where , . If, for some , on no neighbourhood of the origin is of the form for constants , then . Thus , and on .
Interpretation. Lemma 2 shows that the linear and nonlinear components cannot be uniquely separated at a single step size, and longer rollouts at that same step size do not resolve this ambiguity. Theorem 3 states that a second distinct step size removes this ambiguity under the stated conditions. Thus, using two distinct step sizes resolves the ambiguity between the linear and nonlinear components, establishing structural identifiability of the model.
Related work
We review existing methods that support variable-step prediction. Equation-learning methods can recover continuous-time governing equations from data [2] (Rudy et al. 2017; Long et al. 2018). These equations permit flexible time stepping, although prediction still relies on numerical integration of the learned dynamics. Here we focus on operator-learning methods, which learn evolution maps directly. One such approach treats time as a query coordinate, as in DeepONet with a space–time trunk [Lu et al., 2021] and nonlinear coordinate decoders [Seidman et al., 2022]. These models directly learn spatiotemporal neural representations and may require substantial paired training data [Wang et al., 2021]. The second approach, exemplified by Poseidon [6] and OSG-FNO [3], takes the step size as an input. Such models can in principle be evaluated at any step size, but their dependence on is entirely learned through training. Outside the range of observed step sizes—in particular below the finest step the data provide—fitting the available data alone cannot guarantee reliable prediction. A third approach learns a vector field in function space and predicts by numerical integration: TI-DeepONet [Nayak and Goswami, 2026] fits it through a numerical integrator, while CFO [7] fits it to estimated time derivatives. The learned dynamics may consequently inherit errors from the numerical integration or derivative estimation used during training.
In our numerical experiments we compare against one representative of each of the three families above: NOMAD [Seidman et al., 2022] for temporal-spatial representation, OSG-FNO [3] for step-size learning, and TI-DeepONet [Nayak and Goswami, 2026] for learned vector fields. We further compare against E-WSINDy [4] from the equation-learning literature, KNO [Xiong et al., 2024] as a fixed-step method, and persistence.
Numerical experiments
We present four numerical experiments: two synthetic examples and two applications to real-world data. We compare our method with 5 benchmarks discussed in § 2.4. We use the recommended or default settings from the original papers and implementations whenever available and applicable. Otherwise, we select suitable hyperparameters based on validation performance. The architecture, spectral cutoff, symbol degree and data settings of ENNO in each experiment are listed in Appendix § 7.§ 2.1. All neural models are trained for at most 3,000 epochs. All baselines are configured to a parameter count comparable to ours, except OSG-FNO, which we could not reduce below three times ours without seriously degrading its training.
The KdV equation
The first synthetic example is the Korteweg–de Vries equation on the periodic domain . Its symbol lies in the polynomial class of Eq. (6) once , with and all other coefficients zero, so the learned symbol can be compared with the truth coefficient by coefficient. Training uses 2030 trajectories on an grid, corrupted by Gaussian noise at 5% of each trajectory’s own standard deviation; 50 and 200 further trajectories with unseen initial conditions serve for validation and test. The model is trained with two step sizes, and . KNO needs ten frames of context, so all methods are scored on the same targets, ; the single-frame methods are conditioned on and KNO on the ten frames up to and including .
The degree decides whether the true model can be recovered, so we first ablate it in Table 1, with everything else held fixed. The error falls by a factor of 2.7 from to and does not improve clearly at . We use in what follows. We also ablate the latent-field structure, which brings no improvement here, as expected since the true model lies in our class at this degree (Appendix § 7.§ 2.2). To test the identifiability of the two components, we show the recovered coefficients in Table 2, where all of them agree well with the truth. Table 3 compares against the baselines, reporting the mean over the 200 test trajectories with its standard error of the mean (SE); ENNO is lowest at all four target times. The column evaluates each model at half the smallest training step, which none of them has seen: KNO cannot be evaluated there at all, and OSG-FNO and TI-DeepONet are worse at than at despite the shorter horizon.
| 2.895 | 3.676 | 4.200 | |
| 2.179 | 2.695 | 2.989 | |
| 0.313 | 0.721 | 1.115 | |
| 0.319 | 0.743 | 1.151 |
Table 1. KdV: relative error (%) at three target frames, averaged over the 200 test trajectories, for symbol degrees .
| truth | 0 | 0 | 0 | |
| recovered |
Table 2. KdV: recovered symbol coefficients at against the truth.
| Method | ||||
| ENNO () | ||||
| TI-DeepONet | ||||
| E-WSINDy | ||||
| OSG-FNO | ||||
| KNO | — | |||
| NOMAD |
Table 3. KdV: relative error (%) at four target times, reported as mean ± SE over the 200 test trajectories. is half the smallest training step and is not an observation time; methods that cannot be evaluated below the observation interval are marked “—”.
The advection-diffusion equation
The second synthetic example is the advection-diffusion equation with on the periodic domain . Its symbol lies in the polynomial class of Eq. (6) once , with and all other coefficients zero, so the learned symbol can again be compared with the truth coefficient by coefficient. Training uses 100 trajectories on a grid, corrupted by Gaussian noise at 5% of each trajectory’s own standard deviation; 10 and 25 further trajectories serve for validation and test. The encoder and decoder are identity maps, and the model is trained with two step sizes, and . As on KdV, all methods use the last input frame only, except KNO which uses the preceding ten.
The ablation in Table 4 selects , the true order, with no clear improvement at . We first estimate the symbol coefficients: the diffusion coefficient is recovered as 0.0488 against the true 0.05, and the remaining coefficients, whose true values are zero, stay below 0.07 in magnitude (Table 5). Table 6 compares against the baselines at comparable parameter counts. ENNO is lowest at all four target times, with E-WSINDy the closest competitor, unlike on KdV, where its candidate library must cover a third-order dispersive term. At it remains lowest, while KNO cannot be evaluated there at all. Neither branch–trunk method tracks the dynamics here: the error of NOMAD is already at the scale of the solution at and barely changes over the horizon, while that of TI-DeepONet grows from about 15% at to 51% at . Since TI-DeepONet is competitive on KdV (second-best at ), the failure here is not one of training budget.
| 2.439 | 4.011 | 4.647 | |
| 0.021 | 0.074 | 0.129 | |
| 0.021 | 0.074 | 0.128 |
Table 4. Advection–diffusion: relative error (%) at three target frames, averaged over the 25 test trajectories, for symbol degrees .
| truth | 0 | 0/0 | 0.05 |
| recovered | 0.029 | 0.054/0.061 | 0.0488 |
Table 5. Advection–diffusion: recovered symbol coefficients at against the truth. The two first-order coefficients are listed as .
| Method | ||||
| ENNO () | ||||
| E-WSINDy | ||||
| KNO | — | |||
| OSG-FNO | ||||
| TI-DeepONet | ||||
| NOMAD |
Table 6. Advection–diffusion: relative error (%) at four target times, reported as mean ± SE over the 25 test trajectories. is half the smallest training step and is not an observation time; methods that cannot be evaluated below the observation interval are marked “—”.
Radar nowcasting
Radar nowcasting aims to predict the precipitation field 0–2 hours ahead from recent radar observations, filling the gap that numerical weather prediction leaves at these lead times. We take the vertically integrated liquid (VIL) product of SEVIR [Veillette et al., 2020] as the observed field , sampled on a uniform grid over a square domain of km, with consecutive frames five minutes apart. Given the ten most recent frames, the task is to predict the next ten frames at five-minute intervals, covering lead times of 5–50 minutes. We use 11,159 training events, 200 validation events and 500 test events; the test events come from separate files covering November–December 2019. Our model encodes the ten input frames into 20 latent fields and is trained with two step sizes, and minutes. All methods predict the same final ten frames of each test event.
Our model achieves the lowest MSE at all three lead times in Table 7. The comparison with KNO is particularly informative because both models receive the same ten-frame history. Relative to KNO, our model reduces MSE by about 22%, 19% and 11% at 5, 25 and 50 minutes, respectively. The improvement is therefore present in the first prediction and persists at all three lead times, demonstrating a sustained advantage over the tested forecast horizon. Figure 1 illustrates the rollout on a test event. ENNO tracks the motion and coarse structure of the storm cells throughout, while the high-intensity cores are progressively smoothed.
Figure 1. Qualitative SEVIR forecasts for one test event at lead times of 5, 25 and 50 minutes, matching Table 7. Top row: ground truth; bottom row: ENNO’s autoregressive predictions.
| Method | 5 minutes | 25 minutes | 50 minutes |
| ENNO | 1.4 ± 0.05 | 5.4 ± 0.15 | 9.2 ± 0.25 |
| KNO | 1.8 ± 0.05 | 6.7 ± 0.17 | 10.3 ± 0.28 |
| OSG-FNO | 1.8 ± 0.04 | 8.6 ± 0.15 | 13.7 ± 0.26 |
| NOMAD | 5.6 ± 0.14 | 8.9 ± 0.24 | 13.4 ± 0.40 |
| TI-DeepONet | 6.4 ± 0.16 | 11.4 ± 0.29 | 19.5 ± 0.50 |
| E-WSINDy | 12.5 ± 0.30 | Failed | Failed |
| Persistence | 2.6 ± 0.09 | 15.3 ± 0.42 | 24.3 ± 0.66 |
Table 7. SEVIR test MSE () at three forecast lead times, reported as mean SE over the 500 test events. Bold marks the best result. “Failed” denotes nonfinite predictions during rollout.
Wildfire prediction
Wildfire management relies on knowing where a fire will burn next and how intensely. The goal of this example is to predict the next day’s fire radiative power (FRP) field from the FRP observed on the preceding days together with environmental drivers—wind, terrain, vegetation and moisture—recorded on the most recent day. We use WildfireSpreadTS [5], which records 607 wildfire events across the western United States between 2018 and 2021 as daily multi-channel images at 375 m resolution, one channel holding the observed FRP and the rest holding environmental drivers for the same day. We take daily FRP fields as the state, transformed by since FRP spans about four orders of magnitude, and 25 environmental channels from the most recent input day. The latter enter the model as the externally given fields of Eq. (7), so that can modulate the predicted intensity pointwise according to local wind, slope and fuel conditions. Our model is trained with two step sizes, and days; evaluation is always at one day ahead. All neural predictors are trained with the same weighted squared loss on log-FRP, giving weight 50 to pixels with positive FRP and weight 1 to the remaining valid pixels: fire pixels make up under 0.15% of the valid pixels, and without this weighting training can favour near-zero predictions. Pixels introduced by padding events to a common grid size are excluded from both training and evaluation.
We use four folds over 2018–2021, one per test year. Each fold trains on two years, selects a checkpoint on a third, and evaluates on the remaining year. Model selection and evaluation use the area under the precision–recall curve (AUC-PR), with predicted log-FRP serving as a continuous score for detecting pixels with positive FRP. Table 8 shows that our model achieves the highest mean AUC-PR at both history lengths. With one input day, it leads in every held-out year, attaining a mean of 0.3210 compared with 0.2895 for KNO. With five input days, our model and OSG-FNO achieve similar means (0.3122 versus 0.3098), each leading in two of the four held-out years; OSG-FNO uses more than ten times as many trainable parameters (8.4M versus 0.75M). Longer histories do not improve our model’s mean AUC-PR here, and persistence exceeds three of the trained baselines with one input day.
| Method | ||
| ENNO | 0.3210 ± 0.0848 | 0.3122 ± 0.0835 |
| KNO | 0.2895 ± 0.0867 | 0.2889 ± 0.0744 |
| OSG-FNO | 0.2664 ± 0.0811 | 0.3098 ± 0.0887 |
| TI-DeepONet | 0.1435 ± 0.0373 | Unsupported |
| NOMAD | 0.1290 ± 0.0394 | Unsupported |
| E-WSINDy | 0.0337 ± 0.0635 | 0.2533 ± 0.0954 |
| Persistence | 0.2473 ± 0.0984 | 0.2530 ± 0.0972 |
Table 8. WildfireSpreadTS test AUC-PR, reported as mean ± sample standard deviation across four held-out years. Bold marks the highest mean. “Unsupported” indicates that the evaluated implementation does not support five-frame inputs.
Conclusions
We presented ENNO, an operator-learning method that combines exact linear propagation with a local nonlinear correction and prescribes its step-size dependence through a first-order Duhamel approximation. Under suitable conditions, prediction maps at two distinct step sizes locally identify these components, and agreement between models at those steps controls prediction differences across a prescribed range. Experiments on two synthetic PDEs and two real-world datasets demonstrate strong predictive performance, with a latent extension handling the real-data applications. The synthetic examples also show accurate recovery of the principal differential coefficients. The current guarantees concern the base model at fixed spatial resolution. Extending the analysis to the latent construction and obtaining stability bounds uniform under spatial refinement are left to future work.
AI use statement
In this work, we used generative AI tools for tasks that require disclosure as follows. The theoretical results of this paper—the model construction, the statements of Theorem 1, Lemma 2 and Theorem 3, and the proof strategies—were conceived and developed in full by the authors, and complete proofs were written before any AI tool was involved. We then used generative AI tools to check these proofs. The tools identified a number of errors and proposed corrections, some of which we adopted; they also suggested improvements to the arguments, some of which we adopted. Every adopted correction and improvement was verified by the authors before being incorporated, and none of them changed the statements of the results. We disclose this use under both assisting in the writing of proofs and providing critical ingredients for proving mathematical claims.
We have reviewed all AI-assisted work. All AI assistance was carried out under the direction of the authors, and the resulting material was subsequently checked and verified by two authors: every proof, including every AI-suggested correction, was verified line by line independently of the tool’s output; AI-edited text was checked against the claims the authors intended to make. We take responsibility for the final content of this work, including text, claims or artifacts produced with the aid of generative AI.
Reproducibility statement
The model architecture, Fourier cutoff, symbol degree, nonlinear-net inputs and the constraint parametrisation used in each experiment are listed in Appendix § 7.§ 2.1, together with the data-generation procedure for the two synthetic problems and the splits for all four experiments. SEVIR and WildfireSpreadTS are publicly available; the preprocessing applied to each is described in § 3.3 and § 3.4 and Appendix § 7.§ 2.1. Code for the model, the baselines and the evaluation pipeline will be made publicly available.
Appendix
Proofs
Proofs of § 2.3.1
Throughout this appendix, write
For vectors, denotes the norm; for linear operators, it denotes the induced operator norm. Since is finite-dimensional, the locally Lipschitz maps and are Lipschitz on every bounded ball. All Fourier multipliers considered below preserve real-valued fields and commute with one another.
Proof of the estimate in Theorem 1. Set
Then . We may choose
We first control the difference between the linear operators. For and , define
Commutativity of and gives the exact identity
Consequently,
where .
Fix and set . For ,
The same bound holds with replaced by or . The definition of therefore implies
Subtracting Eq. (20) at the two observed step sizes eliminates and yields
Since and , Eq. (13) gives
To control the variation with the step size, define
For , this equals . For , the exponential integral identity gives
For and ,
Thus the Lipschitz bound applies to the nonlinear terms in Eq. (19). It follows that, for ,
The term cancels in this difference.
Now fix and . Set . By Eq. (17),
Moreover,
so . Choose to be the nearer of the two observed step sizes. Equations (21) and (24) imply
Write
The distance to the nearer endpoint satisfies
Therefore,
Taking the supremum over proves Eq. (14), with
All constants have the dependence stated in the theorem. □
We next verify the nonlinear separation condition. The following lemma includes the scalar condition without derivative features.
Lemma 4 (Derivative-free nonlinearities). Suppose , where is locally Lipschitz. Let
Assume that, for some and with ,
Then Eq. (13) holds for some and . These constants can be chosen depending only on .
Proof. Since is continuous and injective on , it is strictly monotone there. Hence there exists such that
Let
Finite dimensionality gives a constant such that for . Choose sufficiently small that
Next choose such that
Fix with . For every real with , the fields
belong to and take values in . Also,
so take values in .
For either of these fields, put . Because and , orthogonality of and Eq. (26) imply
Consequently,
Let . Choose a real unit vector attaining . Since ,
Finally, gives
and hence . Combining these estimates yields
Thus Eq. (13) holds with .
The next lemma covers the quadratic transport nonlinearities in the KdV and advection–diffusion examples.
Lemma 5 (Quadratic transport nonlinearities). Let , with the Euclidean Fourier cutoff and . Suppose
Then, for every , there exist such that Eq. (13) holds.
Proof. Write . Since is homogeneous of degree two,
Its derivative satisfies
We first show that
Suppose the left-hand side holds. Polarization of Eq. (27) gives
This bilinear identity extends to the complexification of . Let denote the Fourier symbol of , and set . Testing the identity on the Fourier modes indexed by and gives
Choose with . Taking is allowed because , and gives . Taking , then gives . For every retained mode with , taking , gives
It remains to consider a nonzero retained mode satisfying . If for some , choose . Then
Thus and are retained and have nonzero -frequency. Equation (28) with therefore implies .
Otherwise, for every . Choose with and . Then , , and . Set . Now
The same use of Eq. (28) gives . Hence for every retained mode, proving .
By finite dimensionality and compactness, this injectivity implies
Homogeneity in and therefore gives
To pass to finite perturbations, introduce the continuous symmetric bilinear map
Choose such that . For , set . The bounds
and the quadratic identity give
Consequently,
Together with Eq. (29), this yields
Taking
proves the claim.
For the KdV equation
take
For the two-dimensional equation
on the periodic domain , take
In both cases, Lemma 5 verifies Eq. (13) for . The linear operators are bounded on , and the projected quadratic nonlinearities are locally Lipschitz. Thus both equations provide reference models satisfying the hypotheses of Theorem 1. The constants may depend on the fixed spatial resolution .
Finally, we establish the sharpness assertion in Theorem 1. We show that the linear growth in
cannot, in general, be replaced by sublinear growth with constants having the dependence stated in the theorem. Take the reference model
For , the scalar contrast is . On every positive interval ,
Lemma 4 therefore verifies the nonlinear separation condition.
Fix the reference model, the parameters in the theorem, and the resulting constants and . Choose
and define a fixed comparison model by
This is a model of the same form, with a locally Lipschitz scalar nonlinearity. A direct calculation gives
where
Both coefficients vanish at , so the two models agree exactly at that step size.
Now take observed steps
and let , keeping both models fixed. The functions and are continuously differentiable on , and is bounded on . Consequently, for a constant independent of ,
Consider the fixed inference step . Since ,
Testing the constant fields yields
On the other hand,
If the amplification factor in the theorem could be replaced by with , then Eq. (31) would make its right-hand side tend to zero:
This contradicts Eq. (32). All model parameters and permitted constant dependencies remain fixed as .
The same argument with the fixed inference step establishes the corresponding sharpness for extrapolation below the observed interval. Thus the linear growth order is optimal in general; no optimality of the numerical coefficient 2 is asserted.
Proofs of Section 2.3.2
Write and . Throughout this section, denotes the original one-step map
In particular, there is no projection of the final nonlinear output in this definition. We write the Fourier modes as on the normalised torus. On a periodic box with side lengths , replace by in the differential symbols. This invertible rescaling preserves the polynomial degrees and the arguments below; the bandwidth bounds refer to integer mode indices.
Proof of Lemma 2. Let . Since , the corresponding linear operator is , and hence
Substitution into (33) gives
where the last equality uses . This proves (16). Repeated composition also gives equality of all rollouts for which both iterates are defined.
For the identifiability theorem, the following explicit choice of bandwidth suffices:
We first establish the two properties supplied by this threshold.
Lemma 6 (Uniform realisation of local features). If , then is surjective for every . Moreover, there are linear right inverses satisfying
where the operator norms use the Euclidean norm on the feature space and the norm on .
Proof. Suppose a real linear functional annihilates the range of . Writing its coefficients as , this means
Complexifying this real linear identity and testing Fourier modes gives
The retained frequencies contain the tensor grid . A polynomial of degree at most in each coordinate that vanishes on this grid is identically zero, by repeated univariate interpolation. Thus all vanish, proving surjectivity. Choose a linear right inverse at one point. Translating its output produces right inverses at all other points, because differentiation commutes with translations. Translations preserve the norm, so these right inverses have the same operator norm.
The next lemma supplies the finite-frequency rigidity needed to identify the differential symbol.
Lemma 7 (Finite-frequency rigidity). Let , , be complex polynomials in real variables, with degrees at most , , , respectively, and suppose . Assume that, for every with ,
If with given by Eq. (34), then is constant.
Proof. We first prove a univariate statement. Let , , be complex polynomials of degrees at most , , , with and nonzero. Suppose that at least distinct real points satisfy
We show that is constant.
Write for real , where and are real polynomials. Taking squared moduli in Eq. (35) gives zeros of
Here and are real polynomials in of degrees at most .
Suppose first that is nonconstant. Differentiating exactly times eliminates the second polynomial and gives
The polynomial is nonzero: at each differentiation, the polynomial factor is transformed by , which maps every nonzero polynomial to a nonzero polynomial when is a nonzero polynomial. Repeated Rolle’s theorem therefore bounds the number of distinct real zeros of by
This contradicts the definition of . Hence is constant. The polynomial then has more zeros than its degree, so
Cancel the common polynomial factors in and write the result as , with and coprime. For a polynomial , let . Equation (36) implies the polynomial identity
In particular, . Coprimality implies that divides , and therefore
The polynomial has no real roots, since any such root would also be a root of . Consequently, the reduced rational function admits a smooth argument on the real line:
If the roots of are , counted with multiplicity, then , , and
In particular,
At most of the sample points are zeros of . At each remaining point, Eq. (35) gives
The range of has width at most , while . Thus at most distinct integers can occur in the preceding identity. By the definition of , one integer occurs at least times. The function
therefore has at least that many distinct zeros.
By Eq. (37), is a rational function with a strictly positive real denominator of degree at most and a numerator of degree at most . Rolle’s theorem supplies at least distinct zeros of this numerator, so it vanishes identically. Hence on . Since as , the polynomial must be zero. Both and are constant, proving the univariate statement.
We now return to the multivariate polynomials. The identity at implies . For each nonzero direction , restrict the polynomials to
The polynomials and are nonzero because their values at zero are nonzero. For every integer with ,
Thus these restrictions satisfy the univariate hypotheses at distinct points. It follows that is constant in , and in particular .
Therefore vanishes on the tensor grid . Its degree in each coordinate is at most , so repeated univariate interpolation gives .
Proof of Theorem 3. We first prove part (i). Choose as in its nondegeneracy assumption and set . Differentiate the equality of the prediction maps at the zero input and evaluate at . For every real , this gives
Define the polynomials
Complexifying the linear identity and testing Fourier modes yields
The nondegeneracy assumption states precisely that
Apply Lemma 7 to
Indeed, Eq. (38) gives , and the two anti-aliasing bounds imply
Hence is constant. Since both differential symbols have real coefficients, their values at zero are real. Thus, for some ,
Equality of the positive-order polynomial coefficients follows immediately.
It remains to identify the nonlinear function on a common feature neighbourhood. Let and be as in Lemma 6. For any step size at which the prediction maps agree on the stated input ball, define
This is positive because the exponential is invertible on the finite-dimensional space . For , take
Then , and
Equality of the two prediction maps at therefore gives
The radius is independent of . Since , this identity holds for every . Taking and proves on , completing part (i).
For part (ii), put
Equation (39) applies at both observed step sizes. Thus, on a common neighbourhood of the origin and for every ,
Suppose , and choose as in the non-affinity assumption. Since ,
Setting in Eq. (40) gives
for all sufficiently close to zero. Define
Then locally. Writing , this implies, on a sufficiently small ball,
Iteration and continuity at zero yield
Consequently,
on a neighbourhood of the origin. This contradicts the assumed non-affinity. Therefore . Part (i) now gives and on its original feature domain .
Numerical Experiments
Main ENNO hyperparameters
Tables 9 and 10 list, for each of the four experiments, the architecture of ENNO and the data it was trained and evaluated on. Throughout, is the Euclidean cutoff on integer mode indices, , as in § 2.3.1; is the degree of the symbol in Eq. (6); is the order of the jet entering the nonlinear map in Eq. (8); and denotes the linearly propagated field to which is applied in Eq. (9). The coefficients of the even monomials in (§ 2.2) are parametrised as , so that holds by construction; the odd-order coefficients and are unconstrained. One consequence is that a constrained coefficient whose true value is zero cannot reach zero exactly and settles at the floor of the parametrisation; this is the case for in Table 2. The unconstrained coefficients are initialised from . The nonlinear map and, where they are not the identity, the encoder and the decoder are multilayer perceptrons with Tanh activations; the output layer of is zero-initialised, so that training starts from the purely linear flow.
ENNO is trained with Adam for at most 3,000 epochs (the cap applied to all neural models in § 3), and its checkpoint is selected by early stopping on the validation set (patience 15 for the synthetic problems, evaluated on the noised validation trajectories, and 200 for the real data; for WildfireSpreadTS the criterion is validation AUC-PR, § 3.4). All reported results come from a single training seed per configuration (per fold for WildfireSpreadTS). Relative errors on the synthetic problems are computed against the clean, noise-free test trajectories; the targets at come from the same solver runs at the finer output interval (Table 10).
| KdV | Advection–diffusion | SEVIR | WildfireSpreadTS | |
| Domain and grid | periodic, 128 points | periodic, | ; VIL on pixels, average-pooled to | daily fire maps at 375 m, padded to |
| Source | Fourier pseudospectral in space, DOP853 in time (rtol ); 10,001 frames over , every 10th kept | explicit Euler, , central differences in space; 5,001 frames over , every 10th kept | real radar, 49 frames per event | real observations, 2018–2021 |
| Initial conditions | 3 sine modes (amplitude for mode ) plus 1 or 2 solitons (amplitude ) | sums of 1–3 Gaussian bumps (amplitude , ) and low-mode Fourier fields | — | — |
| Observation interval | 5 min | 1 day | ||
| Training step sizes | , | , | 5, 10 min | 1, 2 days |
| Targets at | every 5th frame of the same runs | every 5th frame of the same runs | — | — |
| Train / val / test | 2,030 / 50 / 200 trajectories | 100 / 10 / 25 trajectories | 11,159 / 200 / 500 events | four leave-one-year-out folds |
| Noise (train and val, all methods) | i.i.d. Gaussian, 5% of each trajectory’s std, fixed seed; test trajectories clean | same | none | none |
Table 10. Data used in the four experiments.
| KdV | Advection–diffusion | SEVIR (VIL) | WildfireSpreadTS | |
| Latent channels | 1 | 1 | 20 | 4 () / 20 () |
| History length (frames) | 1 | 1 | 10 | |
| Encoder / decoder | identity (learnable variants in Table 11) | identity | pointwise MLPs and , 5 hidden layers 256 | same design, and , 5 hidden layers 256 |
| Cutoff ()1 | 64 | 64 | 64 | 32 |
| Symbol degree | 3 (ablation over 1–4, Table 1) | 2 (ablation over 1–3, Table 4) | 6 | 6 |
| Constant term | trained | trained | fixed at 0 | fixed at 0 |
| Jet order | 2 | 2 | 1 | 1 |
| Inputs of | , , | , , , , , | , , | , , and 25 environmental channels |
| Size of | 4 hidden layers 128 |
Table 9. Architecture of ENNO in the four experiments.
Ablation results of the KdV example
The encoder and decoder of Eq. (12) are identity maps in the configuration reported above, so the model acts directly on . Table 11 makes them learnable, with latent fields, at each symbol degree. Where the degree is below the true order, the learnable variant with is lower at every target time, and the gain saturates between and : at it reaches 2.449% at against 4.189% for the identity. At , where the true symbol is already in the class, the pattern splits: the identity is lower at and , by a factor of 2.7 at , while the learnable variant grows more slowly and ends slightly lower at . We report the identity configuration, which is the more accurate of the two over most of the horizon.
| Variant | |||
| , identity | 2.898 | 3.683 | 4.189 |
| , | 3.069 | 2.629 | 3.276 |
| , | 1.866 | 2.317 | 2.468 |
| , | 1.877 | 2.289 | 2.449 |
| , identity | 2.181 | 2.700 | 2.994 |
| , | 1.990 | 2.454 | 2.856 |
| , | 1.532 | 2.233 | 2.529 |
| , | 1.535 | 2.235 | 2.529 |
| , identity | 0.313 | 0.721 | 1.115 |
| , | 0.852 | 0.915 | 0.994 |
| , | 0.846 | 0.939 | 1.028 |
| , | 0.842 | 0.935 | 1.014 |
| , identity | 0.318 | 0.738 | 1.152 |
| , | 0.846 | 0.920 | 1.007 |
| , | 0.840 | 0.968 | 1.086 |
| , | 0.832 | 0.963 | 1.098 |
Table 11. KdV: encoder/decoder ablation. Relative error (%) at three target frames. “Identity” is the parameter-free configuration reported in the main text; the other rows make them learnable, with latent fields. Best in each block in bold.
References
- [1]Steven L. Brunton and J. Nathan Kutz. Data-Driven Science and Engineering: Machine Learning, Dynamical Systems, and Control. Cambridge University Press, 2019.DOI
- [2]Steven L. Brunton, Joshua L. Proctor, and J. Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 113(15):3932–3937, 2016.DOI
- [3]Junfeng Chen, Kailiang Wu, and Dongbin Xiu. DUE: A deep learning framework and library for modeling unknown equations. SIAM Review, 67(4):873–902, 2025.arxiv.org/abs/2504.10373
- [4]Urban Fasel, J. Nathan Kutz, Bingni W. Brunton, and Steven L. Brunton. Ensemble-SINDy: Robust sparse model discovery in the low-data, high-noise limit, with active learning and control. Proceedings of the Royal Society A, 478(2260):20210904, 2022.DOI
- [5]Sebastian Gerard, Yu Zhao, and Josephine Sullivan. WildfireSpreadTS: A dataset of multi-modal time series for wildfire spread prediction. In Advances in Neural Information Processing Systems, volume 36, 2023.DOI
- [6]Maximilian Herde, Bogdan Raonić, Tobias Rohner, Roger Käppeli, Roberto Molinaro, Emmanuel de Bézenac, and Siddhartha Mishra. Poseidon: Efficient foundation models for PDEs. In Advances in Neural Information Processing Systems, volume 37, pages 72525–72624, 2024.arxiv.org/abs/2405.19101
- [7]Xianglong Hou, Xinquan Huang, and Paris Perdikaris. CFO: Learning continuous-time PDE dynamics via flow-matched neural operators. In International Conference on Learning Representations, 2026.arxiv.org/abs/2512.05297
- [8]Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizdazenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to PDEs. Journal of Machine Learning Research, 24(89):1–97, 2023.
- [9]Jingye Li, Alex Bespalov, and Jinglai Li. Nemytskii neural operator: A nonlinear model reduction method for parametrized partial differential equations. arXiv preprint, 2025. arXiv:2511.07684.arxiv.org/abs/2511.07684
