Equation-Aware Emulation in Surrogate Modeling

• Equation-aware emulation is a design principle that integrates explicit governing equations into surrogate models, enabling zero-shot generalization across unseen PDEs and ODEs.
• It employs methodologies such as PDE-conditioned neural operators, FiLM-style conditioning, and boundary-aware Gaussian processes to dynamically encode and exploit mathematical structure.
• Validated across fluid dynamics, quantum systems, and mechanistic models, this approach demonstrates improved stability, robust control, and enhanced computational efficiency.

Equation-aware emulation denotes a class of surrogate and control methodologies in which explicit equation structure is part of the emulator itself rather than an external interpretation layer. In one prominent formulation, a single neural operator is conditioned on a vector encoding of the governing PDE, so that the equation is an input “prompt” to the emulator rather than a fixed assumption built into its weights (Zhu et al., 12 Nov 2025). In other formulations, the emulator is derived from a target master equation, from a linearized system of governing ODEs, or from analytically known boundaries of a simulator, so that posterior means, variances, or control waveforms are computed directly from the underlying equations (Vlachos et al., 2022, Albert, 2011, Vernon et al., 2018). This suggests that the term refers less to a single architecture than to a shared design principle: expose governing equations, operator structure, or exact analytical substructure to the emulation procedure.

1. PDE-conditioned emulators and the meaning of “equation-aware”

In the PDE-surrogate setting, equation-aware emulation is operationalized as learning a one-step map

$$\hat{u}(t+\Delta t) = \mathcal{M}\big(u(t), \boldsymbol{c}\big),$$

where $u(x,t)$ is the current state and $\boldsymbol{c}\in\mathbb{R}^7$ is a compact encoding of the entire PDE (Zhu et al., 12 Nov 2025). The encoding is defined in a fixed operator basis: $$u_t = c_1 u + c_2 u^2 + c_3 u_x + c_4 u u_x + c_5 u_{xx} + c_6 u_{xxx} + c_7 u_{xxxx},$$ with coefficients for linear reaction, nonlinear reaction, linear advection, nonlinear advection, diffusion, dispersion, and hyper-diffusion. The model is then rolled out autoregressively,

$u^{(k+1)} = \mathcal{M}(u^{(k)}, \boldsymbol{c}).$

This differs from conventional PDE surrogates in three ways. First, the model is parameterized by the PDE itself, not only by initial conditions or a few scalar physical parameters. Second, a single model is trained on multiple structurally distinct 1D PDEs simultaneously. Third, evaluation includes both unseen parameter settings and an entirely unseen PDE whose encoding is supplied only at inference time (Zhu et al., 12 Nov 2025).

The training family in the reported study consists of Korteweg–de Vries, conserved Kuramoto–Sivashinsky, Fisher’s equation, and Advection–Diffusion, all on periodic 1D domains $[0,L]$ with $L=1.0$, 160 grid points, and $\Delta t=1$. Burgers’ equation is completely held out and used only at test time. Its encoding,

$$\boldsymbol{c}_{\text{Burgers}} = [0,0,0,b,\nu,0,0],$$

recombines nonlinear advection $u u_x$ with stabilizing diffusion $u(x,t)$0, testing whether the emulator has learned operator semantics rather than memorized equation identities (Zhu et al., 12 Nov 2025).

A common misconception is to identify equation-aware emulation with parameter-conditioned surrogates for a single PDE family. In the reported framework, however, “generalize to unseen PDEs” means that the model receives an encoding for Burgers’ equation even though no Burgers trajectories were present in training. The task is therefore zero-shot equation-level generalization, not only interpolation in coefficient space (Zhu et al., 12 Nov 2025).

2. Conditioning mechanisms and model architectures

All four reported architectures share the same task definition but differ in how the equation encoding modulates the operator. A central mechanism is FiLM-style conditioning: $u(x,t)$1 where an MLP maps $u(x,t)$2 to channel-wise scale and shift parameters (Zhu et al., 12 Nov 2025).

Model	Core operator	Equation conditioning
PI-FNO-UNET (M1)	U-Net + FNO	FiLM + dynamic spectral weights
LSC-FNO (M2)	Latent neural operator	spectral gating + global attention + FiLM
PINO (M3)	FiLMed FNO	FiLM + encoding projection + PDE-residual loss
LC (M4)	FiLMed ResNet with spectral layers	FiLM on residual correction network

PI-FNO-UNET uses seven input channels,

$u(x,t)$3

with derivatives precomputed by finite differences. Its spectral convolution kernel is generated as a function of $u(x,t)$4, so equation structure modulates both local features and Fourier-space operators. LSC-FNO is a “pure learned features” variant that uses only the state $u(x,t)$5 as input and injects PDE information through spectral gating, global attention, and FiLM. PINO tiles a learned projection of $u(x,t)$6 over space and adds an explicit PDE-residual loss. LC predicts a correction to a coarse numerical solver,

$u(x,t)$7

with $u(x,t)$8 produced by the coarse solver and $u(x,t)$9 learned by a FiLMed residual network (Zhu et al., 12 Nov 2025).

The physics-informed loss in PINO computes the right-hand side implied by the encoding,

$\boldsymbol{c}\in\mathbb{R}^7$0

and penalizes the residual

$\boldsymbol{c}\in\mathbb{R}^7$1

through

$\boldsymbol{c}\in\mathbb{R}^7$2

The total loss is

$\boldsymbol{c}\in\mathbb{R}^7$3

with $\boldsymbol{c}\in\mathbb{R}^7$4 increasing during training up to $\boldsymbol{c}\in\mathbb{R}^7$5 (Zhu et al., 12 Nov 2025).

Training data are generated with APEBench. For each PDE and parameter set, there are 50 training trajectories of 50 time steps and 30 test trajectories of 200 steps. All four training PDEs are mixed into a single corpus, while Burgers is completely held out. The main data loss is MAE over space, or over 5-step unrolled sequences for PINO. Training runs for 100k steps with batch sizes 64–128 and peak learning rates roughly $\boldsymbol{c}\in\mathbb{R}^7$6 (Zhu et al., 12 Nov 2025).

3. Generalization, stability, and zero-shot equation transfer

The principal empirical claims concern stability and generalization under both parameter shift and equation shift. Evaluation uses autoregressive 200-step rollouts, although training trajectories are only 50 steps long. Stability is measured with normalized RMSE,

$\boldsymbol{c}\in\mathbb{R}^7$7

and with an aggregated score

$\boldsymbol{c}\in\mathbb{R}^7$8

using $\boldsymbol{c}\in\mathbb{R}^7$9 for KdV and cKS parameter scans (Zhu et al., 12 Nov 2025).

Within the training PDE family, the models show low nRMSE for both in-distribution and out-of-distribution parameter settings and remain stable over 200 steps. Qualitative rollouts reportedly preserve soliton interactions for KdV and accurate fine-scale chaotic structures for cKS. More significantly, Burgers’ equation is never seen during training; its coefficient vector is merely provided at test time. All four generalized models maintain stable, low-error nRMSE over 200 steps on Burgers, whereas specialized APEBench baselines trained on single PDEs perform poorly and become unstable (Zhu et al., 12 Nov 2025).

The Burgers experiment is especially informative because it tests sign semantics as well as operator reuse. Burgers combines $$u_t = c_1 u + c_2 u^2 + c_3 u_x + c_4 u u_x + c_5 u_{xx} + c_6 u_{xxx} + c_7 u_{xxxx},$$0, seen in KdV and cKS, with positive diffusion $$u_t = c_1 u + c_2 u^2 + c_3 u_x + c_4 u u_x + c_5 u_{xx} + c_6 u_{xxx} + c_7 u_{xxxx},$$1, seen in Fisher and Advection–Diffusion, while cKS uses negative diffusion. Reported zero-shot success therefore indicates that the models distinguish “viscosity as stabilization” from “anti-diffusion as instability” through the sign of $$u_t = c_1 u + c_2 u^2 + c_3 u_x + c_4 u u_x + c_5 u_{xx} + c_6 u_{xxx} + c_7 u_{xxxx},$$2 in $$u_t = c_1 u + c_2 u^2 + c_3 u_x + c_4 u u_x + c_5 u_{xx} + c_6 u_{xxx} + c_7 u_{xxxx},$$3 (Zhu et al., 12 Nov 2025).

Among the four architectures, LC is reported as best overall, with the lowest error and highest robustness; PINO and LSC-FNO also perform very strongly, while PI-FNO-UNET performs slightly worse than the other three. This ranking is interpreted in the study as evidence that correction modeling and latent feature conditioning can be more effective than heavily hand-engineered derivative inputs (Zhu et al., 12 Nov 2025).

These results motivate the paper’s phrase “equation-prompted emulators.” A plausible implication is that explicit operator encodings can move neural operators from per-equation surrogates toward reusable families of surrogates defined over a restricted operator class, provided the target PDE lies within the span of the chosen basis.

4. Mechanistic, GP, and boundary-conditioned formulations

Equation-aware emulation also appears in analytically derived surrogates whose prior or posterior is built from governing equations rather than from purely statistical covariance assumptions. In a mechanistic dynamic emulator for nonlinear ODE simulators, the governing system

$$u_t = c_1 u + c_2 u^2 + c_3 u_x + c_4 u u_x + c_5 u_{xx} + c_6 u_{xxx} + c_7 u_{xxxx},$$4

is linearized as

$$u_t = c_1 u + c_2 u^2 + c_3 u_x + c_4 u u_x + c_5 u_{xx} + c_6 u_{xxx} + c_7 u_{xxxx},$$5

with Gaussian white noise added to account for the nonlinearities; conditioning on $$u_t = c_1 u + c_2 u^2 + c_3 u_x + c_4 u u_x + c_5 u_{xx} + c_6 u_{xxx} + c_7 u_{xxxx},$$6 precomputed nonlinear solutions yields a Gaussian emulator whose online complexity scales like $$u_t = c_1 u + c_2 u^2 + c_3 u_x + c_4 u u_x + c_5 u_{xx} + c_6 u_{xxx} + c_7 u_{xxxx},$$7 in multiplications of matrices of the dimension of the state space, where $$u_t = c_1 u + c_2 u^2 + c_3 u_x + c_4 u u_x + c_5 u_{xx} + c_6 u_{xxx} + c_7 u_{xxxx},$$8 is the number of time-points to emulate (Albert, 2011).

In GP emulation with known boundaries, the relevant equation-awareness is not in the transition law but in exact analytical submanifolds of the simulator. For a hyperplane boundary $$u_t = c_1 u + c_2 u^2 + c_3 u_x + c_4 u u_x + c_5 u_{xx} + c_6 u_{xxx} + c_7 u_{xxxx},$$9 where the simulator is known, Bayesian updating gives

$u^{(k+1)} = \mathcal{M}(u^{(k)}, \boldsymbol{c}).$0

and

$u^{(k+1)} = \mathcal{M}(u^{(k)}, \boldsymbol{c}).$1

where $u^{(k+1)} = \mathcal{M}(u^{(k)}, \boldsymbol{c}).$2 is the projection of $u^{(k+1)} = \mathcal{M}(u^{(k)}, \boldsymbol{c}).$3 onto the boundary and $u^{(k+1)} = \mathcal{M}(u^{(k)}, \boldsymbol{c}).$4 is the perpendicular distance. With two perpendicular boundaries, the variance becomes

$u^{(k+1)} = \mathcal{M}(u^{(k)}, \boldsymbol{c}).$5

The boundary is therefore enforced exactly, uncertainty vanishes on it, and posterior structure remains analytic (Vernon et al., 2018).

A related but more recent dynamic formulation uses one-step-ahead GP emulation together with linked GPs. For ODE-governed simulators, each state component is modeled as

$u^{(k+1)} = \mathcal{M}(u^{(k)}, \boldsymbol{c}).$6

and the paper derives exact analytical posterior means and variances under Gaussian input uncertainty, eliminating the Monte Carlo approximations used in earlier one-step approaches. Numerical experiments on Lotka–Volterra and Lorenz show nearly identical RMSE and CRPS to Monte Carlo-based propagation, but with computation times of about $u^{(k+1)} = \mathcal{M}(u^{(k)}, \boldsymbol{c}).$7 s versus $u^{(k+1)} = \mathcal{M}(u^{(k)}, \boldsymbol{c}).$8 s for Lotka–Volterra and $u^{(k+1)} = \mathcal{M}(u^{(k)}, \boldsymbol{c}).$9 s versus $[0,L]$0 s for Lorenz; the method is implemented in the \textsf{R} package \textsf{dynemu} (Heo, 26 Mar 2025).

Taken together, these formulations show a second major meaning of equation-aware emulation: the emulator is not merely conditioned on data from an equation-governed system, but derives its predictive law, covariance reduction, or propagation algebra from the equation itself.

5. Target-equation emulation in open quantum systems

In open quantum dynamics, equation-aware emulation is used in a literal design sense: a control process is engineered so that the ensemble-averaged dynamics of a physical qubit obey a specified master equation. The target generalized Markovian master equation is

$[0,L]$1

with $[0,L]$2, $[0,L]$3, and $[0,L]$4. The emulation strategy applies a stochastic Hamiltonian

$[0,L]$5

on top of the Markovian Lindblad dynamics and matches the noise autocorrelation to the desired memory kernel,

$[0,L]$6

so that the averaged qubit evolution reproduces the target master equation within stated approximations (Vlachos et al., 2022).

Two concrete kernels are studied: an exponentially decaying memory kernel and a modulated decaying memory kernel. The paper gives constructive routes from the target kernel to the actual control waveform, including random telegraph noise and a Wiener–Khintchine spectral construction. The resulting framework therefore begins from the target equation and synthesizes a classical stochastic drive whose statistics realize it (Vlachos et al., 2022).

This usage of equation-aware emulation is physically different from PDE neural operators or GP surrogates. The goal is not to approximate a simulator faster, but to realize prescribed reduced dynamics on hardware. Nonetheless, the methodological pattern is parallel: equation structure is the design variable. The work also reports that non-Markovian noise can act as a resource for coherence preservation. For Type I noise and $[0,L]$7, the coherence time approaches $[0,L]$8; for Type II, under the stated parameter regime, the upper bound approaches $[0,L]$9. The paper explicitly notes a breakdown regime: when the generalized-Markovian noise is too strong, the decorrelation condition

$L=1.0$0

fails, and the experiment is no longer faithfully described by the target master equation even though stochastic Schrödinger simulations still match the data (Vlachos et al., 2022).

6. Adjacent research areas, misconceptions, and open problems

Several adjacent literatures address problems that are not themselves equation-aware emulation in the narrow surrogate sense, but provide representations, priors, or search mechanisms that can support it. Neural-guided equation discovery with MGMT uses grammar-guided Monte-Carlo Tree Search, dataset embeddings, and policy/value networks; across its experiments, supervised learning outperforms reinforcement learning for almost all module combinations, grammar rules are preferable to tokens as the action space, and risk-seeking MCTS plus AmEx-MCTS improve search efficiency (Brugger et al., 21 Mar 2025). Knowledge-aware equation discovery with EPDE modifies structure-space probabilities rather than imposing rigid templates, using term preference distributions extracted from SymNet or expert knowledge, and is reported to outperform SINDy in search stability and robustness on Burgers, wave, and Korteweg–De Vries examples (Ivanchik et al., 2024). “Informed Equation Learning” represents surrogate hypotheses as compositions of atomic functions with structured sparsity priors and can learn compact analytic models for engineering tasks (Werner et al., 2021). “Equation Embeddings” and EqEmb-U learn semantic vectors for equations from surrounding words and equation units, improving pseudo log-likelihood over CBOW, GloVe, PV-DM, and Bernoulli embeddings across four arXiv domains (Krstovski et al., 2018). Similarity-based equational inference treats derivations as sequences of symbolic state transitions and reconstructs intermediate equations with heuristic search over CAS-level actions (Meadows et al., 2021).

These adjacent systems clarify a second misconception: equation-awareness is not identical to symbolic regression. Symbolic discovery reconstructs equations from data; semantic embeddings organize equations in representation space; derivation inference reconstructs missing algebraic steps. Emulation, by contrast, targets predictive substitution for a simulator or target dynamics. A plausible implication is that future equation-aware emulators may combine all three layers: explicit equation prompts or operator bases, symbolic discovery of governing structure, and learned representations of equation similarity.

Reported limitations are substantial. In the PDE-neural-operator setting, current evidence is restricted to 1D periodic PDEs, a seven-term operator basis, and APEBench’s default settings; extending to 2D or 3D, other boundary conditions, nonlocal operators, or variable coefficients requires broader encodings and higher-dimensional architectures (Zhu et al., 12 Nov 2025). Mechanistic dynamic emulators rely on linearization plus Gaussian noise, so strong nonlinearities or regime changes may not be captured by a single global approximation (Albert, 2011). Linked GP dynamic emulators rely on Gaussian moment propagation and can degrade on long chaotic horizons, even though they remain much faster than Monte Carlo propagation (Heo, 26 Mar 2025). Master-equation emulation depends on weak-coupling and decorrelation assumptions and fails in regimes of sufficiently strong generalized-Markovian noise (Vlachos et al., 2022).

Across these usages, the durable idea is that emulators improve when equations are not treated as hidden causes of data but as explicit computational objects. Whether that object is a PDE coefficient vector, a Lindbladian with memory kernel, a linearized ODE operator, or a known analytic boundary, equation-aware emulation turns governing structure into an active part of inference, rollout, and uncertainty quantification.