Learning missing physics from legacy simulators with alternating neural integrators Authors: Hao Wang, Qinghe Wang, Caiyou Yuan, Kailiang Wu Published: 2026 ABSTRACT A recurring challenge in science and engineering is the model–reality gap, where trusted legacy simulators lose fidelity due to unresolved physics or structural incompleteness. This challenge has motivated remedies ranging from imperfect mechanistic models to fully data-driven surrogates. Here, we address this gap with Alternating Neural Integrators (ANI), a non-intrusive reuse-and-correct framework for upgrading executable legacy simulators without requiring access to or modification of their internal implementation. Guided by operator-splitting principles, ANI alternates the evolution of a fixed prior simulator with a learned neural correction that targets structured discrepancy between the prior and the supervisory data. We show that ANI can recover missing coupling in chaotic systems and act as an effective subgrid correction in turbulence, improving dynamical fidelity where prior models drift. In selected mechanistically structured settings, post hoc symbolic distillation yields compact hypotheses and, in controlled benchmarks, supports closed-loop refinement of the prior. By combining data-driven flexibility with reusable scientific simulators in a theoretically grounded framework, this work provides a practical gray-box route for systematically upgrading existing computational infrastructure when a callable prior and supervisory data are available. Introduction From forecasting climate change to designing next-generation aircraft and discovering new materials, computational models are indispensable across science and engineering. Yet even trusted simulators can lose predictive fidelity when unresolved physics, reduced-order closures, coarse discretizations, or simplifying assumptions become consequential, creating a systematic model–reality gap between mechanistic predictions and the dynamics ultimately observed. In many scientific settings, the challenge is therefore not the absence of a model, but the presence of one that is informative, executable, and yet structurally incomplete. When experimental observations or high-fidelity simulations reveal this discrepancy, a practical tension often arises: whether to continue relying on an imperfect mechanistic model or to replace it with a fully data-driven surrogate. The central question, however, is how an existing, trusted simulator can be upgraded non-intrusively, without rewriting it from the ground up. In this work, we use the term missing physics in an operational sense to denote the structured discrepancy between a reusable prior and the reference dynamics, rather than to imply that every learned correction corresponds to a new fundamental law. Scientific machine learning has developed a broad repertoire of hybrid strategies to address this gap, including discrepancy or gray-box modeling1, semigroup-aware and operator-splitting-inspired learning of evolution operators2,3, closure modeling traditions in large-eddy simulation and turbulence correction4–6, physics-informed learning7,8, operator learning9–11, and pretraining/fine-tuning strategies for surrogate models12. While learning residual or discrepancy structure is not new, a practical limitation remains in many settings: many existing frameworks require access to an explicit, differentiable formulation of the prior. The specific contribution of this work is to retain a callable prior step map inside each corrected update through alternating composition. This distinction is important for legacy simulators, compiled solvers, reduced models, and frozen neural priors, where intrusive reformulation or automatic differentiation may be impractical. Here we introduce alternating neural integrators (ANI), a non-intrusive prior–corrector framework for upgrading reusable but incomplete simulators by alternating a trusted prior step with a learned neural correction that targets structured discrepancy between the prior and the supervisory data. This operator-splitting construction supports even-order extensions, admits mathematically grounded error analysis through a modified residual flow in the time-discrete setting, and applies across a broad range of executable priors. Across ten benchmark ordinary and partial differential equation systems, together with two scoped real-data case studies (single-cell battery aging and single-basin rainfall–runoff forecasting), ANI improves long-horizon accuracy and dynamical fidelity relative to the prior alone and to matched learned baselines; in representative ODE/PDE settings, it also outperforms matched additive gray-box baselines. In selected mechanistically structured cases, post hoc symbolic distillation yields compact hypotheses and, in controlled benchmarks, supports closed-loop refinement of the prior. Taken together, this work makes three contributions. First, it formulates a non-intrusive alternating-composition framework that wraps callable prior step maps rather than requiring explicit differentiable vector fields. Second, it provides mathematically grounded analysis under standard regularity conditions for the resulting modified residual flow in the time-discrete setting. Third, it evaluates the framework across benchmark systems, scoped real-data case studies, and selected post hoc write-back examples. Figure 1 provides a schematic overview of the ANI prior–corrector framework. In practical terms, ANI makes established scientific codebases reusable components within modern machine-learning workflows without requiring refactoring of their internal implementation. Its theoretically grounded design, supported by a modified residual-flow analysis, provides error and stability results under standard regularity assumptions (smooth local flows on compact reachable sets with bounded derivatives up to the order required by the expansion) and helps explain the observed long-horizon robustness. ANI is therefore aimed at the operating regime in which reusable prior and supervisory data coexist, changing what can be learned from that combination. Results We evaluate ANI on ten benchmark ordinary and partial differential equation problems together with a single-cell battery-aging study and a single-basin rainfall–runoff study. These studies span chaotic systems, oscillators, excitable media, compressible flow with shock-containing profiles, and turbulence, as well as battery aging13 and observational rainfall–runoff forecasting on CAMELS-US14. The priors cover empirical laws, coarse or low-order solvers, reduced-order models, incomplete physical models, established numerical schemes, coarse-grid approximations, and learned surrogates, such as Sparse Identification of Nonlinear Dynamics (SINDy)15,16, Kolmogorov–Arnold Networks (KAN)17,18, and frozen neural propagators. As summarized in Table 1, across the benchmark suite and the two scoped real-data studies, ANI reduces error growth and improves long-horizon stability and dynamical diagnostics relative to the prior alone and to matched learned baselines. For the benchmark ordinary and partial differential equation tasks, high-fidelity numerical or exact solutions serve as references, whereas for the real-data studies we train and evaluate against observed trajectories paired with frozen mechanistic priors.Table 1ANI across ten benchmark ordinary and partial differential equation problemsExperimentPrior modelCore challengeCharacteristic dynamicsKey finding of ANI1. Damped pendulumUndamped pendulum (missing dissipation)Capture damping effects on amplitude and phaseRecovers the missing dissipative component and corrects amplitude and phase.2. Lorenz–StenfloLorenz subsystem (structurally incomplete)Restore long-horizon chaotic dynamics with sensitive dependenceRestores the attractor and chaos diagnostics where the prior collapses.3. Glycolytic oscillatorSINDy-identified affine model (approximate)Model strong nonlinear oscillations and phase shiftsRefines the symbolic prior and captures the nonlinear oscillatory regime.4. Damped compound pendulumUndamped double pendulum (missing dissipation)Capture multi-frequency damping and couplingRecovers coupled damping effects in a more intricate mechanical system.5. Morris–LecarCoarse solver (forward Euler)Recover fine dynamics lost by coarse integrationRecovers fast spiking features beyond the coarse solver.6. Multi-trophic predator–preyKAN-based prior (black-box)Model multi-species interactions and cyclesImproves population-cycle prediction over the initial neural prior.7. 1D EulerTVD finite differences (numerical diffusion)Model shock-containing profiles while controlling spurious oscillationsSharpens shock-associated gradients and contacts relative to the stabilized TVD prior.8. FitzHugh–NagumoFrozen neural prior (fixed regime)Generalization under parameter shiftBroadens the operating range of a frozen neural prior.9. Navier–Stokes (2D)Partial physics (viscous term only)Reconstruct full dynamics from an incomplete priorLearns the missing transport and restores flow statistics.10. Kolmogorov flowLarge-eddy simulation (coarse grid)Learn an effective subgrid correctionRestores spectra and finescale structures as an effective subgrid correction.The table summarizes the benchmark suite used to evaluate ANI across structurally deficient physical models, coarse numerical solvers, reduced-order closures, and approximate learned priors. Two additional scoped real-data studies (single-cell battery aging and single-basin CAMELS-US rainfall–runoff forecasting) are reported separately in the Results. SINDy sparse identification of nonlinear dynamics, KAN Kolmogorov–Arnold network, TVD total variation diminishing, and LES large-eddy simulation. Table 1ANI across ten benchmark ordinary and partial differential equation problemsExperimentPrior modelCore challengeCharacteristic dynamicsKey finding of ANI1. Damped pendulumUndamped pendulum (missing dissipation)Capture damping effects on amplitude and phaseRecovers the missing dissipative component and corrects amplitude and phase.2. Lorenz–StenfloLorenz subsystem (structurally incomplete)Restore long-horizon chaotic dynamics with sensitive dependenceRestores the attractor and chaos diagnostics where the prior collapses.3. Glycolytic oscillatorSINDy-identified affine model (approximate)Model strong nonlinear oscillations and phase shiftsRefines the symbolic prior and captures the nonlinear oscillatory regime.4. Damped compound pendulumUndamped double pendulum (missing dissipation)Capture multi-frequency damping and couplingRecovers coupled damping effects in a more intricate mechanical system.5. Morris–LecarCoarse solver (forward Euler)Recover fine dynamics lost by coarse integrationRecovers fast spiking features beyond the coarse solver.6. Multi-trophic predator–preyKAN-based prior (black-box)Model multi-species interactions and cyclesImproves population-cycle prediction over the initial neural prior.7. 1D EulerTVD finite differences (numerical diffusion)Model shock-containing profiles while controlling spurious oscillationsSharpens shock-associated gradients and contacts relative to the stabilized TVD prior.8. FitzHugh–NagumoFrozen neural prior (fixed regime)Generalization under parameter shiftBroadens the operating range of a frozen neural prior.9. Navier–Stokes (2D)Partial physics (viscous term only)Reconstruct full dynamics from an incomplete priorLearns the missing transport and restores flow statistics.10. Kolmogorov flowLarge-eddy simulation (coarse grid)Learn an effective subgrid correctionRestores spectra and finescale structures as an effective subgrid correction.The table summarizes the benchmark suite used to evaluate ANI across structurally deficient physical models, coarse numerical solvers, reduced-order closures, and approximate learned priors. Two additional scoped real-data studies (single-cell battery aging and single-basin CAMELS-US rainfall–runoff forecasting) are reported separately in the Results. SINDy sparse identification of nonlinear dynamics, KAN Kolmogorov–Arnold network, TVD total variation diminishing, and LES large-eddy simulation. ANI across ten benchmark ordinary and partial differential equation problems The table summarizes the benchmark suite used to evaluate ANI across structurally deficient physical models, coarse numerical solvers, reduced-order closures, and approximate learned priors. Two additional scoped real-data studies (single-cell battery aging and single-basin CAMELS-US rainfall–runoff forecasting) are reported separately in the Results. SINDy sparse identification of nonlinear dynamics, KAN Kolmogorov–Arnold network, TVD total variation diminishing, and LES large-eddy simulation. Across all tasks, the comparison protocol is matched as closely as the problem class permits. ANI, the frozen prior, and the reported learned baselines are trained and evaluated on identical downstream splits and target trajectories. When a purely learned baseline is shown, it is capacity-matched to the ANI corrector backbone or, in PDE settings, uses the same operator family as the frozen prior where appropriate. The additive gray-box baselines reuse the same prior and residual ingredients and differ only in the coupling rule. For the real-data studies, mechanistic priors are calibrated on the training portion and then frozen before downstream comparison. Exact preprocessing, architecture, and split details are reported in the Supplementary Information. We highlight four benchmark case studies in the Results and then turn to two real-data case studies that probe applicability in experimental and observational regimes. The benchmark cases probe complementary capabilities: restoring missing coupling in a chaotic system, learning an effective subgrid correction in turbulence, broadening a frozen neural prior under parameter shift, and recovering a known dissipative correction in a controlled damped mechanical example. Extended visualizations, additional gray-box comparisons, cost/accuracy sweeps, pretraining controls, and symbolic-distillation details are reported in the Supplementary Information. Matched additive gray-box comparisons are included for both the Lorenz–Stenflo and FitzHugh–Nagumo benchmarks. A compact summary of representative quantitative outcomes for these case studies is provided in Supplementary Table 7, while the key comparisons supporting the main claims are reported in the corresponding subsections below. Even when the priors are structurally flawed, they retain inductive bias that ANI can exploit by learning only the discrepancy relative to the reference evolution. Recovering missing coupling in a chaotic system We study the Lorenz–Stenflo equations 19 as a controlled chaotic benchmark in which the prior model is structurally incomplete. The reference dynamics are four-dimensional, whereas the prior evolves the Lorenz subsystem in (x, y, z) while carrying w as a frozen auxiliary coordinate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\dot{w}=0)$$\end{document}(w°=0), thereby omitting the coupling through w. In this setting, the task is not to make a small parametric adjustment to an otherwise correct model, but to learn an effective correction that compensates for a missing coupling mechanism and restores the correct long-horizon dynamics. This benchmark is challenging for single-stage learned residual baselines. An additive gray-box RK4 baseline, built from the same prior and a learned residual, provides only limited short-horizon accuracy but does not recover the correct attractor geometry or tangent-space instability statistics. ANI-4, by contrast, improves all three reported chaotic diagnostics relative to the baseline. Supplementary Table 11 reports mean ± SEM over four trajectories on fixed 500-step evaluation windows, and Supplementary Fig. 16 shows the corresponding diagnostic-window robustness. Across longer windows, the baseline remains systematically low in approximate entropy and correlation dimension, whereas ANI-4 tracks the reference more closely. ANI-2 already improves correlation dimension and largest Lyapunov exponent relative to the baseline, while ANI-4 gives the closest overall agreement across all three diagnostics. In particular, the largest Lyapunov-exponent error drops from 20.9% for the baseline to 2.1% for ANI-4. The baseline’s poor short-term accuracy (Fig. 2a) and the prior’s collapse to a fixed point (Fig. 2c) underscore the magnitude of the challenge. In contrast, the ANI-4 prediction (Fig. 2e) closely matches the reference attractor (Fig. 2b). In this controlled benchmark, ANI yields a corrected hybrid model from a structurally deficient prior and, through post hoc distillation, a compact correction hypothesis for the omitted coupling mechanism. Supplementary Section 9.2 further reinserts this distilled correction into the prior equations and shows that the resulting symbolic refined model reduces the mean 200-step relative L2 error from 4.6 × 10−1 for the frozen prior to 5.7 × 10−3 while retaining a much sparser support than the additive discrepancy alternative. A complementary symbolic write-back comparison against an additive-SINDy distillation further shows that the ANI-4 distilled law more faithfully preserves the attractor geometry and chaos diagnostics after compression (Supplementary Table 15 and Supplementary Fig. 18).Fig. 1Alternating neural integrators (ANI) upgrade an executable prior simulator with a learned residual corrector.a A reusable but imperfect prior model (Prior-only, dashed green) can deviate from the reference dynamics (solid blue) because unresolved or mismodeled effects are not fully captured. b ANI treats the prior simulator as a callable black-box component and alternates it with a compact learned corrector that models an effective discrepancy relative to the prior and the supervisory data. The same construction supports higher-order variants built from the base update. c The corrected rollout (dashed orange) tracks the reference dynamics more faithfully than the prior alone, illustrating how ANI upgrades an existing simulator without requiring access to or modification of its internal implementation.Fig. 2Comparison of physical prior, additive gray-box baseline, ANI-2, and ANI-4 on the Lorenz–Stenflo benchmark.a Relative L2 error versus time. b Reference attractor. c Prior-only trajectory produced by the frozen Lorenz subsystem. d Additive gray-box baseline built from the same prior plus a learned residual integrated with fourth-order Runge–Kutta (RK4). e ANI-4 using the same prior and residual ingredients but coupled by alternating composition. Alternating neural integrators (ANI) upgrade an executable prior simulator with a learned residual corrector. a A reusable but imperfect prior model (Prior-only, dashed green) can deviate from the reference dynamics (solid blue) because unresolved or mismodeled effects are not fully captured. b ANI treats the prior simulator as a callable black-box component and alternates it with a compact learned corrector that models an effective discrepancy relative to the prior and the supervisory data. The same construction supports higher-order variants built from the base update. c The corrected rollout (dashed orange) tracks the reference dynamics more faithfully than the prior alone, illustrating how ANI upgrades an existing simulator without requiring access to or modification of its internal implementation. Comparison of physical prior, additive gray-box baseline, ANI-2, and ANI-4 on the Lorenz–Stenflo benchmark. a Relative L2 error versus time. b Reference attractor. c Prior-only trajectory produced by the frozen Lorenz subsystem. d Additive gray-box baseline built from the same prior plus a learned residual integrated with fourth-order Runge–Kutta (RK4). e ANI-4 using the same prior and residual ingredients but coupled by alternating composition. ANI-4 is more accurate at early and intermediate times because the higher-order composition reduces local truncation error more effectively. At longer horizons in this chaotic regime, however, the error curves of ANI-2 and ANI-4 become closer because both models remain confined to the correct attractor while pointwise trajectory errors saturate near the intrinsic predictability limit. Selected post hoc distillation results from this benchmark are summarized later in Table 2 and Supplementary Section 9.1, where the learned correction is converted into a compact candidate refinement hypothesis for the prior equations.Table 2Post hoc distillation results from ANISystemSupportDistilled dominant correctionValidation statusInterpretationLorenz–Stenflo4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \dot{x}\approx 1.506\,w$$\end{document}Δx°≈1.506w; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \dot{w}\approx -0.035-1.005\,x-0.998\,w$$\end{document}Δw°≈−0.035−1.005x−0.998wClosed-loop validated after reinsertion into the prior. The refined symbolic model achieves mean 200-step relative L2 error 5.7 × 10−3, compared with 4.6 × 10−1 for the prior.Compact correction hypothesis for the omitted coupling in the prior equations.FitzHugh–Nagumo4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \dot{v}\approx 0.701+0.564\,u-0.140\,v+0.142\,uv$$\end{document}Δv°≈0.701+0.564u−0.140v+0.142uvConsistency check against the idealized affine parameter-discrepancy term; the learned correction is relative to an imperfect frozen neural prior and may include prior-error-induced terms. Additional details are provided in Supplementary Section 9.3.Correction retains the dominant affine component expected from the idealized parameter discrepancy, while the small effective interaction term reflects prior error.Battery aging11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \dot{V}\approx 0.8101-0.535V-0.102I+0.585\,{{{\rm{SoC}}}}+0.060\,{{{\rm{Cyc}}}}-0.188V\,{{{\rm{SoC}}}}+\cdots$$\end{document}ΔV°≈0.8101−0.535V−0.102I+0.585SoC+0.060Cyc−0.188VSoC+⋯Empirical refinement hypothesis for the equivalent-circuit prior; the full sparse law is reported in Supplementary Section 9.4.Age- and state-dependent empirical correction hypotheses relative to the frozen equivalent-circuit prior.Full symbolic expressions and additional discussion appear in Supplementary Section 9. Here, SoC denotes the state of charge and Cyc denotes the normalized cycle index. Table 2Post hoc distillation results from ANISystemSupportDistilled dominant correctionValidation statusInterpretationLorenz–Stenflo4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \dot{x}\approx 1.506\,w$$\end{document}Δx°≈1.506w; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \dot{w}\approx -0.035-1.005\,x-0.998\,w$$\end{document}Δw°≈−0.035−1.005x−0.998wClosed-loop validated after reinsertion into the prior. The refined symbolic model achieves mean 200-step relative L2 error 5.7 × 10−3, compared with 4.6 × 10−1 for the prior.Compact correction hypothesis for the omitted coupling in the prior equations.FitzHugh–Nagumo4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \dot{v}\approx 0.701+0.564\,u-0.140\,v+0.142\,uv$$\end{document}Δv°≈0.701+0.564u−0.140v+0.142uvConsistency check against the idealized affine parameter-discrepancy term; the learned correction is relative to an imperfect frozen neural prior and may include prior-error-induced terms. Additional details are provided in Supplementary Section 9.3.Correction retains the dominant affine component expected from the idealized parameter discrepancy, while the small effective interaction term reflects prior error.Battery aging11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \dot{V}\approx 0.8101-0.535V-0.102I+0.585\,{{{\rm{SoC}}}}+0.060\,{{{\rm{Cyc}}}}-0.188V\,{{{\rm{SoC}}}}+\cdots$$\end{document}ΔV°≈0.8101−0.535V−0.102I+0.585SoC+0.060Cyc−0.188VSoC+⋯Empirical refinement hypothesis for the equivalent-circuit prior; the full sparse law is reported in Supplementary Section 9.4.Age- and state-dependent empirical correction hypotheses relative to the frozen equivalent-circuit prior.Full symbolic expressions and additional discussion appear in Supplementary Section 9. Here, SoC denotes the state of charge and Cyc denotes the normalized cycle index. Post hoc distillation results from ANI Full symbolic expressions and additional discussion appear in Supplementary Section 9. Here, SoC denotes the state of charge and Cyc denotes the normalized cycle index. Learning effective subgrid correction in turbulence We next consider coarse-grid large-eddy simulation (LES)4,6 of Kolmogorov flow with forcing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\bf{f}}}}(x,y)=(0.1\,\sin (8\pi y),0)$$\end{document}f(x,y)=(0.1sin(8πy),0) and viscosity ν = 10−4. The prior is a 128 × 128 LES, which inherently filters out small-scale features relative to the 1024 × 1024 reference. In this setting, the model-reality gap corresponds to unresolved subgrid dynamics. Rather than replacing the turbulence solver, ANI learns an effective, data-driven subgrid correction layered on top of the executable LES before recovering these missing multiscale effects. To probe temporal extrapolation, training samples are randomly drawn from the first half of the time horizon, whereas validation and test rollouts are selected exclusively from the second half to ensure strict temporal separation. The learned residual functions as an effective subgrid correction that improves long-horizon fidelity relative to the LES prior without disrupting the resolved cascade structure. As shown by the kinetic energy spectrum (Fig. 3b), both the LES prior and the ANI-enhanced models remain broadly consistent with the expected k−5/3 inertial-range and k−3 enstrophy-cascade scalings over the resolved range, with the sharp high-wavenumber drop in the LES curve reflecting the spectral dealiasing cutoff of the coarse solver. The critical distinction is revealed in the long-horizon trajectory error: ANI-2 and ANI-4 achieve a consistent reduction in relative vorticity error compared to the LES prior alone (Fig. 3a). This improved fidelity comes with an expected cost–accuracy trade-off: as reported in Supplementary Section 10.1, a direct FNO baseline remains the fastest inference-time option because it avoids repeated prior calls. Complementary evidence from a related Navier–Stokes crude-prior benchmark further shows that the benefit of prior-corrected learning persists under limited high-fidelity supervision, including against a pretrain–fine-tune FNO baseline (Supplementary Section 10.3). Figure 3c adds a representative vorticity snapshot at t = 1.0 s: the coarse LES field attenuates some fine-scale structures, whereas ANI restores sharper reference-consistent resolved vortical features. Supplementary Figs. 4, 5 provide additional snapshots and corresponding velocity fields, consistent with an effective non-intrusive subgrid correction rather than a simple denoiser.Fig. 3ANI acts as an effective non-intrusive subgrid correction in this turbulence benchmark.The prior is the coarse-grid large-eddy simulation. a Relative L2 vorticity error versus time shows that both ANI-2 and ANI-4 consistently outperform the LES prior. b The kinetic energy spectrum at t = 1.5 s shows that the LES prior and ANI-enhanced rollouts remain broadly consistent with the expected resolved-range cascade-aligned scalings, while the sharp high-wavenumber drop of the LES curve reflects the spectral dealiasing cutoff of the coarse solver. c A representative vorticity snapshot at t = 1.0 s shows that ANI restores sharper reference-consistent resolved vortical features that are attenuated by the LES prior. Spectra are computed after projecting all fields to the common 128 × 128 resolved grid. ANI acts as an effective non-intrusive subgrid correction in this turbulence benchmark. The prior is the coarse-grid large-eddy simulation. a Relative L2 vorticity error versus time shows that both ANI-2 and ANI-4 consistently outperform the LES prior. b The kinetic energy spectrum at t = 1.5 s shows that the LES prior and ANI-enhanced rollouts remain broadly consistent with the expected resolved-range cascade-aligned scalings, while the sharp high-wavenumber drop of the LES curve reflects the spectral dealiasing cutoff of the coarse solver. c A representative vorticity snapshot at t = 1.0 s shows that ANI restores sharper reference-consistent resolved vortical features that are attenuated by the LES prior. Spectra are computed after projecting all fields to the common 128 × 128 resolved grid. Correcting a neural prior under parameter shift We next consider a setting in which the prior is itself a frozen pretrained neural surrogate for the FitzHugh–Nagumo system. This prior is accurate only in a fixed parameter setting and degrades under shifts in the system parameters (ε, a, b), providing a useful test of whether ANI can correct not only mechanistic solvers but also callable learned propagators. Importantly, ANI does not require an explicit continuous-time formula for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{A}}}}(u)$$\end{document}A(u). It requires only an executable prior step map. In this experiment, the role of the prior is played by a frozen one-step neural propagator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{P}}}}}_{h}^{\eta }$$\end{document}Phη, and ANI composes this map with a separate learned corrector while keeping the prior parameters η fixed. The learned residual should therefore be interpreted here as an effective correction to the frozen prior map at step size h, rather than as a correction to a closed-form continuous-time vector field. Concretely, the frozen prior is a hybrid FNO–ETDRK4 surrogate inspired by the Exponential Time-Differencing fourth-order Runge–Kutta method (ETDRK4)20. The prior employs 8 Fourier modes in each direction, width 40, and 4 layers, while the ANI corrector is a lightweight parameter-conditioned CNN residual module (Supplementary Section 8.3). The results in Fig. 4 show that ANI markedly broadens the operating range of this prior. Figure 4b reports the relative L2 error as a function of ε ∈ [4, 6] (with a = 0.75, b = 0.75), and Fig. 4c shows the errors over (a, b) ∈ [0.6, 0.8] × [0.7, 0.9] (with ε = 5.5). In Fig. 4c, the prior colormap is clipped at 1 for visualization, and different color scales are used across columns to resolve the much lower-error ANI regimes; numerical values should therefore be compared using the corresponding color bars. Each grid cell reports the mean over four independently generated test trajectories. The isolated saturated salt-and-pepper pixels primarily reflect rare sample-dependent rollout instabilities of the frozen prior FNO–ETDRK4 composition in this regime, rather than a physical bifurcation boundary over the sampled parameter range.Fig. 4ANI broadens the operating range of a frozen neural prior under parameter shift.A pretrained neural propagator, accurate only for a narrow parameter range of the FitzHugh–Nagumo system, serves as the prior. a Mean relative error versus rollout time under parameter shift; shaded bands indicate ± 1 standard deviation across test trajectories. b Average relative error as a function of parameter ϵ. c Error heatmaps over the (a, b) parameter space; the three columns in (c) report the frozen prior, ANI-2, and ANI-4, respectively. The frozen prior exhibits large errors over broad parts of the sweep, whereas both ANI variants remain substantially more accurate across the sampled parameter space. Color scales differ across columns to resolve the lower-error ANI regimes; numerical values should be compared using the corresponding colorbars. ANI broadens the operating range of a frozen neural prior under parameter shift. A pretrained neural propagator, accurate only for a narrow parameter range of the FitzHugh–Nagumo system, serves as the prior. a Mean relative error versus rollout time under parameter shift; shaded bands indicate ± 1 standard deviation across test trajectories. b Average relative error as a function of parameter ϵ. c Error heatmaps over the (a, b) parameter space; the three columns in (c) report the frozen prior, ANI-2, and ANI-4, respectively. The frozen prior exhibits large errors over broad parts of the sweep, whereas both ANI variants remain substantially more accurate across the sampled parameter space. Color scales differ across columns to resolve the lower-error ANI regimes; numerical values should be compared using the corresponding colorbars. Away from its fixed reference setting, the frozen prior exhibits rollout errors that frequently exceed 100%. In contrast, ANI-2 and ANI-4 remain substantially more accurate across the parameter sweep, indicating that the learned correction captures how the discrepancy depends on (ε, a, b) rather than merely fitting a static bias. Supplementary Section 12 shows that ANI-4 also outperforms a single-stage additive gray-box baseline on this PDE example, while Supplementary Section 10.1 reports the corresponding inference-time trade-off: direct FNO inference is faster because it avoids repeated prior calls, whereas ANI improves long-horizon accuracy by embedding learned corrections within the frozen prior dynamics. Figure 5 contrasts ANI with a single-stage additive gray-box baseline built from the same frozen prior map and a learned residual increment. ANI-4 maintains lower mean relative L2 error across the ε sweep and accumulates mean relative L2 error more slowly over time, indicating that the advantage is not merely due to adding a correction term but persists against a competitive additive prior-plus-residual alternative. Across the (a, b) sweep, ANI-4 also has consistently lower mean relative L2 error than the matched additive baseline, indicating that the gain is not attributable to residual learning alone but also to the alternating coupling rule. Full parameter-space heatmaps and implementation details are reported in Supplementary Section 12.Fig. 5Compact PDE gray-box comparison on FitzHugh–Nagumo.The additive gray-box baseline uses the same frozen prior step map but applies a single-stage residual correction instead of ANI's alternating composition. Across the ε sweep (a) and over rollout time (b), ANI-4 attains lower mean relative L2 error than this additive prior-plus-residual alternative. Compact PDE gray-box comparison on FitzHugh–Nagumo. The additive gray-box baseline uses the same frozen prior step map but applies a single-stage residual correction instead of ANI's alternating composition. Across the ε sweep (a) and over rollout time (b), ANI-4 attains lower mean relative L2 error than this additive prior-plus-residual alternative. Controlled recovery of missing dissipation As a controlled low-dimensional benchmark, we consider a pendulum whose true dynamics differ from the prior by a single missing damping term. The ground-truth dynamics are 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{{{\rm{d}}}}{{{\bf{u}}}}}{{{{\rm{d}}}}t}=\left[\begin{array}{c}\omega \\ -0.3\,\omega -9.80665\,\sin \theta \end{array}\right]={{{\mathcal{A}}}}({{{\bf{u}}}})+{{{\mathcal{B}}}}({{{\bf{u}}}}),$$\end{document}dudt=ω−0.3ω−9.80665sinθ=A(u)+B(u),where u = (θ, ω)T, while the prior retains only the conservative part 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{A}}}}({{{\bf{u}}}})=\left[\begin{array}{c}\omega \\ -9.80665\,\sin \theta \end{array}\right].$$\end{document}A(u)=ω−9.80665sinθ.In this regime, the missing mechanism is known, so the experiment serves as a controlled mechanistic validation case. The learnable module Φθ is trained to approximate the effective missing dissipative component induced by the chosen split and discretization, which in this example closely corresponds to the continuous damping term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{B}}}}({{{\bf{u}}}})={(0,-0.3\omega )}^{T}$$\end{document}B(u)=(0,−0.3ω)T over the tested regime. Figure 6 shows that ANI remains stable and substantially more accurate than a Neural-RK6 baseline that learns the entire vector field from data alone. The phase portrait demonstrates that the ANI prediction closely matches the ground-truth spiraling trajectory toward the stable equilibrium, indicating that the learned residual isolates an effective dissipative correction with high fidelity in the tested regime. This controlled example also illustrates the benefit of higher-order composition: the error decreases systematically from ANI-2 to ANI-6. We therefore use ANI-4 as the default high-order choice in the more demanding chaotic and turbulent demonstrations.Fig. 6ANI recovers a missing dissipative mechanism in a controlled mechanical system with a conservative Hamiltonian prior.The prior model is a conservative, undamped pendulum. a Total relative L2 error over time for ANI models of increasing order (ANI-2, ANI-4, ANI-6) compared to a Neural-RK6 baseline; higher-order ANI variants reduce the error relative to ANI-2 in this tested regime, with ANI-4 and ANI-6 giving the smallest errors. b Phase portrait of the long-term trajectory. The ANI-6 prediction closely matches the ground-truth spiraling trajectory toward the stable equilibrium, demonstrating recovery of the missing effective dissipation in the tested regime. ANI recovers a missing dissipative mechanism in a controlled mechanical system with a conservative Hamiltonian prior. The prior model is a conservative, undamped pendulum. a Total relative L2 error over time for ANI models of increasing order (ANI-2, ANI-4, ANI-6) compared to a Neural-RK6 baseline; higher-order ANI variants reduce the error relative to ANI-2 in this tested regime, with ANI-4 and ANI-6 giving the smallest errors. b Phase portrait of the long-term trajectory. The ANI-6 prediction closely matches the ground-truth spiraling trajectory toward the stable equilibrium, demonstrating recovery of the missing effective dissipation in the tested regime. We next consider two real-data settings: a single-cell battery-aging study and a single-basin rainfall–runoff study. In both cases, the targets are measured rather than generated by a trusted high-fidelity simulator, while the prior remains a reusable but imperfect mechanistic model. These studies extend ANI to measured-data settings and are presented as focused case studies. Single-cell battery-prior drift on experimental data We consider the NASA PCoE Li-ion battery dataset13 and use cell B0005 as the experimental case study. Cell B0005 contains 168 charge–discharge cycles. We use cycles 0–133 for training, 134–149 for validation, and 150–167 as the held-out aged test set. In this setting, the true electrochemical degradation dynamics are high-dimensional and unknown, while standard equivalent-circuit models remain useful but incomplete. We employ a standard Thevenin-type 1RC equivalent-circuit model as the mechanistic prior. Across the battery analyses, the state-of-charge (SoC) input feature is computed by Coulomb counting using the reference-capacity convention specified for each evaluation protocol; these conventions and the corresponding preprocessing details are summarized in Supplementary Table 2. Numerical comparisons are made within each reported protocol. The prior parameters are fit on the training cycles and then frozen, so the resulting model captures the overall discharge shape but cannot account for aging-induced drift in capacity, impedance, and voltage relaxation. The purely data-driven baseline uses the same one-step inputs and the same multilayer-perceptron backbone as the ANI corrector and differs only in the absence of prior coupling. We compare ANI against both this frozen mechanistic prior and the matched data-driven baseline. Under the nominal-capacity visualization convention used in Fig. 7, the representative held-out aged discharge cycle 150 shows that the frozen 1RC prior exhibits a large systematic bias, the matched data-driven baseline reduces but does not remove that mismatch, and ANI-4 gives the closest reconstruction in this representative-cycle comparison. Separately, under the initial-capacity multicycle rollout protocol over the 18 held-out aged cycles 150–167, ANI improves the aggregate metrics and per-cycle rollout errors relative to both baselines. Aggregate metrics over the 18 held-out aged cycles are reported in Supplementary Table 3, the corresponding per-cycle behavior appears in Supplementary Fig. 1, and protocol details are summarized in Supplementary Table 2.Fig. 7Correcting mechanistic-prior drift on single-cell experimental aging data from the NASA B0005 Li-ion battery.Cycle 150 is shown as a representative held-out aged discharge cycle under the nominal-capacity SoC convention; aggregate multicycle metrics over cycles 150–167 are reported separately in Supplementary Table 3 under the multicycle evaluation convention. a–d compare the frozen mechanistic prior, the matched data-driven baseline, ANI-2, and ANI-4 on this representative aged cycle. Numerical comparisons should be interpreted within each reported protocol, and battery preprocessing details are specified in Supplementary Table 2. Correcting mechanistic-prior drift on single-cell experimental aging data from the NASA B0005 Li-ion battery. Cycle 150 is shown as a representative held-out aged discharge cycle under the nominal-capacity SoC convention; aggregate multicycle metrics over cycles 150–167 are reported separately in Supplementary Table 3 under the multicycle evaluation convention. a–d compare the frozen mechanistic prior, the matched data-driven baseline, ANI-2, and ANI-4 on this representative aged cycle. Numerical comparisons should be interpreted within each reported protocol, and battery preprocessing details are specified in Supplementary Table 2. Under the multicycle rollout protocol, the mean RMSE decreases from 0.163 V for the frozen prior and 0.093 V for the matched data-driven baseline to 0.048 V for ANI-2 and 0.039 V for ANI-4 (Supplementary Table 3). The corresponding cycle-wise traces in Supplementary Fig. 1 confirm that this improvement persists across all 18 held-out aged test cycles. Supplementary Section 2 documents the exact preprocessing, split construction, fairness-controlled MLP architectures, and aged-cycle evaluation protocol. Single-basin rainfall–runoff forecasting on CAMELS-US We report a fixed single-basin case study on CAMELS-US Basin 01031500. The basin was selected before analysis because it combines a nearly complete filtered record with a stable calibrated prior, rather than on the basis of downstream performance. The daily record is split chronologically into train/validation/test segments: 1980-01-01 to 2007-10-18, 2007-10-19 to 2011-04-09, and 2011-04-10 to 2014-09-30, respectively. Exact preprocessing, model settings, basin metadata, and the fixed-basin rationale are documented in Supplementary Tables 4–6. The prior is a frozen linear-reservoir model driven by daily precipitation and temperature, and we compare it against the calibrated prior, a Transformer-only baseline, and ANI hybrids under long-horizon autoregressive rollout. On this fixed single-basin case study, the calibrated prior drifts strongly under autoregressive rollout, whereas the Transformer-only baseline improves NSE/KGE from −0.087/−0.042 to 0.817/0.797 and reduces ∣PBIAS∣ from 82.5% to 10.3%. ANI-4 further improves these metrics to NSE 0.843, KGE 0.825, and ∣PBIAS∣ 8.9%, indicating a modest but consistent gain in this basin (Supplementary Table 6). Figure 8 compares multi-year hydrographs for the calibrated prior, the Transformer-only baseline, ANI-2, and ANI-4 over the held-out test period. In this basin, coupling a learned correction to a simple mechanistic prior improves long-horizon rollout quality.Fig. 8Hydrograph comparison for a single-basin CAMELS-US case study.Multi-year autoregressive streamflow forecasts on Basin 01031500 over the held-out test period (2011-04-10 to 2014-09-30). a–d compare the calibrated prior, Transformer-only baseline, ANI-2, and ANI-4, respectively. The prior drifts under long-horizon rollout, the Transformer baseline improves skill, and the ANI hybrids further reduce peak and recession mismatches over the same test interval. Hydrograph comparison for a single-basin CAMELS-US case study. Multi-year autoregressive streamflow forecasts on Basin 01031500 over the held-out test period (2011-04-10 to 2014-09-30). a–d compare the calibrated prior, Transformer-only baseline, ANI-2, and ANI-4, respectively. The prior drifts under long-horizon rollout, the Transformer baseline improves skill, and the ANI hybrids further reduce peak and recession mismatches over the same test interval. From discrepancy to closed-loop refinement We evaluate predictive performance across all tasks and examine post hoc distillation in selected cases where compact corrections are plausible. The three examples below span distinct evidential levels: closed-loop validated refinement for Lorenz–Stenflo, a controlled consistency check for FitzHugh–Nagumo, and an empirical correction hypothesis on experimental battery data. Table 2 summarizes the distilled outputs and their validation status. For the Lorenz–Stenflo benchmark, the distilled ANI-4 correction captures the dominant omitted coupling with nearly minimal support, and Supplementary Section 9.2 reinserts this law into the prior equations for a closed-loop rerollout. In FitzHugh–Nagumo, the distilled correction retains the dominant affine component expected from the idealized parameter discrepancy, with a small effective interaction induced by imperfections in the frozen neural prior. In the battery study, the distilled correction yields age- and state-dependent empirical correction hypotheses relative to the equivalent-circuit prior. These examples illustrate how, in selected settings, a learned discrepancy can be turned into a candidate mechanistic refinement. Discussion ANI provides a non-intrusive pathway for upgrading reusable prior models by learning structured discrepancies. Its core contribution lies not in discrepancy learning itself—which is already well established in gray-box scientific machine learning—but in a specific alternating-composition framework that retains callable prior step maps at the target and fractional substep sizes used by the selected composition inside each corrected update. In this design, the prior contributes trusted structure at each step, while the learned component focuses on the discrepancy relative to the reference dynamics and supervisory data. Across the benchmark suite, this coupling improves long-horizon prediction and dynamical diagnostics relative to both the frozen prior and matched learned baselines. In representative ODE/PDE settings, matched additive gray-box controls on Lorenz–Stenflo and FitzHugh–Nagumo further suggest that these gains do not arise from residual learning alone but also reflect the alternating coupling rule. The battery and hydrology studies extend ANI to measured-data regimes and complement the controlled benchmarks. Two features help explain these gains. First, the prior supplies inductive bias: even when incomplete, a reusable solver often preserves useful geometry, scales, invariants, or numerical structure that helps constrain the learning problem. Second, alternating composition places the learned correction inside the evolution operator itself rather than attaching it as an outer additive patch. This yields a cleaner separation between the trusted prior contribution and the learned discrepancy, and it is precisely this regime that the modified residual-flow analysis addresses. In the time-discrete setting, that analysis provides assumption-qualified error and stability statements under standard regularity assumptions, consistent with the improved long-horizon robustness observed empirically. ANI is designed for settings in which a usable prior and paired supervisory data are available. Because the learned corrector can depend on discretization, boundary conditions, numerical stabilization, and operating regime, substantial shifts in these factors may require retraining or fine-tuning. The method trades inference speed for stronger prior anchoring and, in the regimes studied here, improved long-horizon fidelity and data efficiency. When inference latency or query throughput is the primary deployment constraint, especially in some PDE applications (with varying initial/boundary conditions or parameters) requiring many surrogate evaluations, ANI-4 may not be the preferred low-latency choice; a direct neural operator surrogate, such as FNO, can be preferable when its accuracy is sufficient for the target application. Consistent with this scope, the measured-data results are shown for a single measured cell and a fixed basin. While algorithmically non-intrusive, ANI also supports practical integration with compiled solvers by minimizing the need for internal code refactoring. The potential of ANI to support hypothesis generation is illustrated by its ability, in selected mechanistically structured settings, to recover interpretable structures from model error. Across the tasks studied here, ANI learns a discrepancy operator that improves prediction. In selected low-dimensional or mechanistically structured settings, this learned operator can be distilled post hoc into compact symbolic hypotheses and, in controlled benchmarks, support a closed-loop path from prior model to learned correction to mechanistic refinement. In more complex systems, however, the neural residual itself may already constitute a scientifically meaningful outcome, even if it does not collapse to a simple closed form. ANI is therefore best understood as a pragmatic gray-box framework for upgrading reusable legacy simulators when callable prior step maps for the required substep sizes and paired supervisory data are available. Under those conditions, structured discrepancy is not only something to be reduced; it can also become an informative object of analysis in its own right. Methods We formalize ANI by decomposing the reference dynamics into a reusable prior operator and an effective residual operator. This decomposition yields alternating high-order compositions around callable prior step maps at the substep sizes used by the selected composition and motivates the assumption-qualified analysis below. The modified residual operator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{{{{\mathcal{B}}}}}}_{2n}$$\end{document}B^2n serves as the analytical handle that connects the conceptual framework to a practical implementation, while the prior evolution is treated as a callable black-box component so that the same framework applies to legacy solvers, coarse numerical pipelines, and frozen learned propagators. Problem formulation and prior-residual decomposition Consider a dynamical system governed (implicitly) by ordinary or time-dependent partial differential equations, 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{{\rm{d}}}u}{{{\rm{d}}}t}\,=\,{{{\mathcal{L}}}}(u),\;\; u(0)={u}_{0},$$\end{document}dudt=L(u),u(0)=u0,where the (possibly nonlinear) operator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{L}}}}$$\end{document}L is unknown. Suppose there already exists an imperfect but usable prior model: 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{{\rm{d}}}u}{{{\rm{d}}}t}\,=\,{{{\mathcal{A}}}}(u),$$\end{document}dudt=A(u),with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{A}}}}$$\end{document}A coming from coarse/low-order numerics, a reduced model (e.g., LES), a known subsystem, or a learned surrogate. We define the residual operator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{B}}}}$$\end{document}B by 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{L}}}}(\,\cdot )\,=\,{{{\mathcal{A}}}}(\,\cdot )+{{{\mathcal{B}}}}(\,\cdot ).$$\end{document}L(⋅)=A(⋅)+B(⋅).This operator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{B}}}}$$\end{document}B is the mathematical representation of the missing physics: the unknown effective discrepancy relative to the prior and the reference dynamics. The exact evolution over step h reads \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u(t+h)=\exp \{h({{{\mathcal{A}}}}+{{{\mathcal{B}}}})\}u(t)$$\end{document}u(t+h)=exp{h(A+B)}u(t). For analysis we write the prior in generator form through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\exp \{h{{{\mathcal{A}}}}\}$$\end{document}exp{hA}, but implementation only requires callable prior step maps \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{P}}}}}_{\tau }$$\end{document}Pτ for the target and fractional substep sizes τ used by the selected composition. We therefore distinguish the formal analytic flow \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\exp \{h{{{\mathcal{A}}}}\}$$\end{document}exp{hA} from an executable prior propagator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{P}}}}}_{h}$$\end{document}Ph, which may be delivered by a legacy solver, a reduced model, or a frozen neural surrogate. ANI builds a high-order surrogate by alternating this callable prior evolution with a trainable residual component. Deployment uses only the executable prior step maps \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{P}}}}}_{\tau }$$\end{document}Pτ required by the selected composition. High-order alternating multi-product expansions We build even-order (2n) schemes from a time-symmetric second-order block (Strang-like21) via multi-product expansions22. The symmetric second-order block has an even-order error expansion, so multi-product extrapolation cancels successive lower-order local error terms and yields a formal local error O(h2n+1). ANI approximates the true flow \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\exp \{h({{{\mathcal{A}}}}+{{{\mathcal{B}}}})\}$$\end{document}exp{h(A+B)} by symmetric alternating sub-evolutions generated by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{A}}}}$$\end{document}A and a trainable residual. For h > 0, the second-order composition is 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Psi }_{2}(h)\,=\,\exp \left\{\frac{h}{2}{{{\mathcal{A}}}}\right\}\,\exp \left\{h\,{{{\mathcal{B}}}}\right\}\,\exp \left\{\frac{h}{2}{{{\mathcal{A}}}}\right\}.$$\end{document}Ψ2(h)=exph2AexphBexph2A.General even-order multi-product alternating schemes read 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Psi }_{2n}(h)\,=\,{\sum}_{i=1}^{n}{c}_{i}{\left({\Psi }_{2}\left(\frac{h}{{k}_{i}}\right)\right)}^{{k}_{i}}\,=\,\exp \left\{h({{{\mathcal{A}}}}+{{{\mathcal{B}}}})\right\}\,+\,{{{\mathcal{O}}}}({h}^{2n+1}).$$\end{document}Ψ2n(h)=∑i=1nciΨ2hkiki=exph(A+B)+O(h2n+1).Here ki are distinct positive integers and ci are the associated Vandermonde coefficients, as listed in Supplementary Section 5.1. The final multi-product update is a pointwise linear combination of branch outputs in the discretized state space, rather than a single sequential composition. We use the standard choice ki = i for i = 1, …, n, which yields, e.g., the fourth-order scheme 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Psi }_{4}(h)\,=\,-\,\frac{1}{3}\,{\Psi }_{2}(h)\,+\,\frac{4}{3}\,{\left[{\Psi }_{2}\left(\frac{h}{2}\right)\right]}^{2}.$$\end{document}Ψ4(h)=−13Ψ2(h)+43Ψ2h22. Because the symmetric second-order block has an even-order error expansion, the multi-product extrapolation cancels successive low-order error terms and yields a formal even-order approximation. A modified residual map as the analytical handle For analysis, the finite-step residual correction can be represented through an ideal branchwise residual map, or equivalently through its formal logarithmic generator. For notational brevity, we write this branchwise generator as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{{{{\mathcal{B}}}}}}_{2n}$$\end{document}B^2n; within the i-th component of the multi-product formula, this shorthand refers to the corresponding generator associated with the substep h/ki, with the precise convention given in Supplementary Remark 5.5. With this branchwise shorthand, the following exact map-level identity holds 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sum}_{i=1}^{n}{c}_{i}{\left(\exp \left\{\frac{h}{2{k}_{i}}{{{\mathcal{A}}}}\right\}\exp \left\{\frac{h}{{k}_{i}}{\widehat{{{{\mathcal{B}}}}}}_{2n}\right\}\exp \left\{\frac{h}{2{k}_{i}}{{{\mathcal{A}}}}\right\}\right)}^{{k}_{i}}\,=\,\exp \left\{h({{{\mathcal{A}}}}+{{{\mathcal{B}}}})\right\},$$\end{document}∑i=1nciexph2kiAexphkiB^2nexph2kiAki=exph(A+B),where ∑ici = 1. This notation is used as an analytical shorthand for the branchwise exact residual maps, rather than as a claim that the multi-product map is generated by a single unique autonomous residual vector field. All regularity and approximation constants below are understood uniformly over the finitely many branchwise modified residual generators used by the selected composition. Precise definitions and proofs are given in Supplementary Section 5.2. This ideal residual map provides an analytical target for the effective residual correction that ANI approximates in practice. It is the key analytical handle that connects our conceptual framework to a stable, high-order, and practical implementation. In shock-containing or noisy settings, this object should be interpreted as an effective correction to the stabilized discrete flow map realized by the executable numerical pipeline rather than as a smooth continuum correction. Prior operator vs. ANI corrector (architecture). ANI decomposes the dynamics into (i) a fixed family of prior step maps \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{{{{{\mathcal{P}}}}}_{\tau }\}$$\end{document}{Pτ}, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{P}}}}}_{\tau }\approx \exp \{\tau {{{\mathcal{A}}}}\}$$\end{document}Pτ≈exp{τA} for the required substep sizes τ, and (ii) a trainable corrector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Phi }_{\theta }(h)\approx \exp \{h{\widehat{{{{\mathcal{B}}}}}}_{2n}\}$$\end{document}Φθ(h)≈exp{hB^2n} that models the modified residual dynamics introduced above. The prior step family \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{{{{{\mathcal{P}}}}}_{\tau }\}$$\end{document}{Pτ} is treated as a non-intrusive black box and is frozen during ANI training; it may be provided by a mechanistic simulator, a numerical scheme, or a pretrained neural prior that can be evaluated at the required substep sizes. The corrector Φθ is the only learned component in ANI and is instantiated to match the state representation: for ODEs we use an MLP, while for PDEs we use a discretized functional operator. The second-order ANI building block uses the symmetric composition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Psi }_{2}^{h}={{{{\mathcal{P}}}}}_{h/2}\circ {\Phi }_{\theta }(h)\circ {{{{\mathcal{P}}}}}_{h/2}$$\end{document}Ψ2h=Ph/2∘Φθ(h)∘Ph/2, and higher-order ANI-2n uses weighted multi-product compositions that call \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{P}}}}}_{\tau }$$\end{document}Pτ at the corresponding fractional substep sizes. Flow-map and vector-field realizations Let Φθ(h) denote a neural module conditioned on step size h. We propose two complementary realizations that share the same alternating scaffold and analysis. (i) Residual flow-map learning. Φθ(h) directly approximates the flow \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\exp \{h\,{\widehat{{{{\mathcal{B}}}}}}_{2n}\}$$\end{document}exp{hB^2n} on compact state sets. The ANI-2n operator is 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{N}}}}}_{2n}(h;\theta )\,=\,{\sum}_{i=1}^{n}{c}_{i}{\left(\exp \left\{\frac{h}{2{k}_{i}}{{{\mathcal{A}}}}\right\}{\Phi }_{\theta }\left(\frac{h}{{k}_{i}}\right)\exp \left\{\frac{h}{2{k}_{i}}{{{\mathcal{A}}}}\right\}\right)}^{{k}_{i}}.$$\end{document}N2n(h;θ)=∑i=1nciexph2kiAΦθhkiexph2kiAki.Architectures for Φθ may be MLPs, convolutional/spectral blocks, or operator-learning variants; universal approximation on compacta ensures representational adequacy (Supplementary Sections 5.4–5.5). (ii) Residual vector-field learning. Alternatively, the network may parameterize a residual vector field \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{{{{\mathcal{B}}}}}}_{\theta }$$\end{document}B^θ, which is advanced by a classical one-step integrator (e.g., Runge–Kutta) to define the residual flow map Φθ( ⋅ ) used in Eq. (10). This realizes an approximation to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\exp \{(\cdot ){\widehat{{{{\mathcal{B}}}}}}_{2n}\}$$\end{document}exp{(⋅)B^2n} and is advantageous when the physics suggests a meaningful decomposition (advection/diffusion, linear/nonlinear, stiff/non-stiff). Variable time-step learning High-order compositions inherently require residual evaluations at fractional sub-steps (e.g., h/2 in ANI-4). We therefore condition Φθ on the step size h, either by concatenating h to inputs or via feature modulation (e.g., FiLM23), thereby approximating a continuous family \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\{{\Phi }_{\theta }(h)\}}_{h\, > \,0}$$\end{document}{Φθ(h)}h>0 and avoiding re-training for each h. This design supports CFL-controlled step sizes (e.g., shock tubes) and improves robustness under multi-resolution rollouts. Training data, losses, and non-intrusive coupling We assemble supervised pairs 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{D}}}}={\left\{({u}_{k}^{{{{\rm{in}}}}},\,{u}_{k}^{{{{\rm{out}}}}},\,{h}_{k})\right\}}_{k=1}^{N},$$\end{document}D=(ukin,ukout,hk)k=1N,where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u}_{k}^{{{{\rm{out}}}}}$$\end{document}ukout is a high-fidelity evolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u}_{k}^{{{{\rm{in}}}}}$$\end{document}ukin over step hk (Supplementary Section 8 details per-task generation). Parameters θ are fitted by minimizing 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{L}}}}(\theta )\,=\,\frac{1}{N}{\sum}_{k=1}^{N}d\,\left({{{{\mathcal{N}}}}}_{2n}({h}_{k};\theta )\,{u}_{k}^{{{{\rm{in}}}}},\,{u}_{k}^{{{{\rm{out}}}}}\right),$$\end{document}L(θ)=1N∑k=1NdN2n(hk;θ)ukin,ukout,with d( ⋅ , ⋅ ) the task-appropriate metric (e.g., L2 for ODE states, Lp for fields; see Supplementary Section 3). Stochastic optimizers, such as AdamW24, are used with standard schedules/regularization. When automatic differentiation through the prior is available, gradients are propagated end-to-end. When the prior is an external compiled black-box solver, we use either a one-sided forward-only coupling or a differentiable proxy for symmetric compositions, while keeping deployment coupled to the compiled solver; see Supplementary Section 13. Multi-horizon supervision can also be added when desired: while Eq. (12) is the basic one-step objective used for benchmark training unless otherwise noted, the two measured-data studies additionally use short-rollout or curriculum losses as specified in Supplementary Section 2; the analysis below covers both single- and multi-step settings via Lipschitz-type perturbation bounds. Implementation and non-intrusive coupling ANI can accommodate diverse discretizations and prior forms, provided that callable prior step maps \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{P}}}}}_{\tau }$$\end{document}Pτ are available for the target and fractional substep sizes used by the chosen composition. For ordinary differential equations we employ shallow multilayer perceptrons with smooth activations; for partial differential equations we use lightweight convolutional or residual blocks with periodic padding when appropriate. The prior step \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\exp \{h{{{\mathcal{A}}}}\}$$\end{document}exp{hA} may be a coarse solver (for example forward Euler for Morris–Lecar), a reduced-order closure (for example large-eddy simulation), a conservative subsystem (for example diffusion-only Navier–Stokes), or a learned surrogate (for example SINDy/KAN wrapped in Runge–Kutta). Supplementary Fig. 2 summarizes this flow-map view. In every case, the prior is called as a black box at the required substep sizes inside Eq. (10), enabling non-intrusive prior–corrector coupling with existing codebases without rewriting the solver internals. Computational cost and performance A single ANI-2n step for state u and step h comprises the following operations.For i = 1, …, n: set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widetilde{h}=h/{k}_{i}$$\end{document}h~=h/ki and define \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{F}}}}}_{i}:={{{{\mathcal{P}}}}}_{\widetilde{h}/2}\circ {\Phi }_{\theta }(\widetilde{h})\circ {{{{\mathcal{P}}}}}_{\widetilde{h}/2}$$\end{document}Fi:=Ph~/2∘Φθ(h~)∘Ph~/2.Compose ki times: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{i}={{{{\mathcal{F}}}}}_{i}^{\,{k}_{i}}(u)$$\end{document}vi=Fiki(u).Return \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{N}}}}}_{2n}(h;\theta )u\,=\,{\sum }_{i=1}^{n}{c}_{i}\,{v}_{i}$$\end{document}N2n(h;θ)u=∑i=1ncivi.Per step, ANI-2n invokes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sum }_{i=1}^{n}2{k}_{i}$$\end{document}∑i=1n2ki prior calls and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sum }_{i=1}^{n}{k}_{i}$$\end{document}∑i=1nki residual evaluations (e.g., ANI-4 uses 3 residual and 6 prior sub-steps, as shown in Supplementary Algorithm 1). In the present PDE examples, ANI attains lower long-horizon error than the direct learned baselines at the rollout steps used here while inheriting useful numerical structure from the prior. For i = 1, …, n: set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widetilde{h}=h/{k}_{i}$$\end{document}h~=h/ki and define \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{F}}}}}_{i}:={{{{\mathcal{P}}}}}_{\widetilde{h}/2}\circ {\Phi }_{\theta }(\widetilde{h})\circ {{{{\mathcal{P}}}}}_{\widetilde{h}/2}$$\end{document}Fi:=Ph~/2∘Φθ(h~)∘Ph~/2. For i = 1, …, n: set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widetilde{h}=h/{k}_{i}$$\end{document}h~=h/ki and define \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{F}}}}}_{i}:={{{{\mathcal{P}}}}}_{\widetilde{h}/2}\circ {\Phi }_{\theta }(\widetilde{h})\circ {{{{\mathcal{P}}}}}_{\widetilde{h}/2}$$\end{document}Fi:=Ph~/2∘Φθ(h~)∘Ph~/2. Compose ki times: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{i}={{{{\mathcal{F}}}}}_{i}^{\,{k}_{i}}(u)$$\end{document}vi=Fiki(u). Compose ki times: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{i}={{{{\mathcal{F}}}}}_{i}^{\,{k}_{i}}(u)$$\end{document}vi=Fiki(u). Return \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{N}}}}}_{2n}(h;\theta )u\,=\,{\sum }_{i=1}^{n}{c}_{i}\,{v}_{i}$$\end{document}N2n(h;θ)u=∑i=1ncivi. Return \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{N}}}}}_{2n}(h;\theta )u\,=\,{\sum }_{i=1}^{n}{c}_{i}\,{v}_{i}$$\end{document}N2n(h;θ)u=∑i=1ncivi. Runtime comparisons. Direct neural operators can be faster at inference because they do not call the prior solver; corresponding FNO timing comparisons for PDE rollouts are reported in Supplementary Section 10.1. ANI trades some of this speed for stronger inductive bias and improved long-horizon fidelity when a reusable prior is available. Empirical runtime measurements for representative FitzHugh–Nagumo and Kolmogorov-flow rollouts are summarized in Supplementary Table 16. We report empirical runtime cost as the total wall-clock time required to roll out all test trajectories in a batch, rather than timing very short single-trajectory runs. Wall-clock times are always measured on identical test sets and under identical hardware conditions. All experiments were run on a single NVIDIA A100 GPU. For the FitzHugh–Nagumo benchmark, the reported timing uses four test trajectories rolled out for 200 steps. The corresponding accuracy curves visualize the first 100 steps, i.e., up to t = 1 s. The FNO baseline directly advances the solution with a learned one-step map and therefore has the lowest inference cost. In contrast, ANI-k decomposes each target step into alternating prior and neural correction stages. All ANI variants use the same target step size, but their runtime increases with the number of prior-plus-neural stages, which explains the higher cost of ANI-4 compared with ANI-2. For the Kolmogorov-flow benchmark, the reported timing uses twenty test trajectories rolled out for 400 steps with target step size h = 0.005 s, corresponding to T = 2 s. The large-eddy simulation prior cannot advance stably with such a large internal time step and therefore integrates each target step using smaller substeps no larger than 5 × 10−4 s. Each ANI-k rollout is then constructed by alternating this prior evolution with neural correction stages over the target step h. As the number of stages increases, more prior evaluations are required, leading to a higher runtime of ANI-4 compared with the prior and ANI-2. These results quantify the expected speed–accuracy trade-off: direct FNO inference is fastest, whereas higher-order ANI compositions incur additional computational cost in exchange for substantially improved rollout fidelity. Evaluation metrics and baselines We quantify trajectory and statistical fidelity via relative Lp errors, Lyapunov exponents, correlation dimension, approximate entropy, and turbulence spectra, with precise definitions collected in Supplementary Section 3. High-fidelity numerical or exact solutions serve as references for the benchmark problems; training data are generated from these references while priors are kept fixed. For the real-data studies, observed trajectories are paired with frozen mechanistic priors. Unless otherwise noted, we compare second- and fourth-order ANI variants against the prior-only solver and capacity-matched learned baselines, including pure learned integrators and, where appropriate, additive gray-box alternatives. Task-specific data generation, architectures, and hyperparameters appear in Supplementary Section 8. Theoretical analysis under standard regularity assumptions The analysis addresses the time-discrete prior-corrector dynamics on compact reachable sets under standard regularity assumptions. In non-smooth or noisy regimes, ANI is best interpreted as learning a discretization-consistent correction to the stabilized numerical flow rather than as a continuum statement. All constants are defined on compact reachable sets. (i) Universal approximation. After spatial discretization, sufficiently expressive neural modules can approximate the target residual maps on compact sets under standard universal-approximation assumptions; in particular, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Phi }_{\theta }(h)\approx \exp \{h\,{\widehat{{{{\mathcal{B}}}}}}_{2n}\}$$\end{document}Φθ(h)≈exp{hB^2n} in the flow-map realization, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{{{{\mathcal{B}}}}}}_{\theta }\approx {\widehat{{{{\mathcal{B}}}}}}_{2n}$$\end{document}B^θ≈B^2n in the vector-field realization, with arbitrarily small approximation error within the adopted approximation class (Supplementary Sections 5.3–5.5). (ii) Single-step ANI error analysis. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{U}}}}(h)=\exp \{h({{{\mathcal{A}}}}+{{{\mathcal{B}}}})\}$$\end{document}U(h)=exp{h(A+B)} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{N}}}}}_{2n}(h;\theta )$$\end{document}N2n(h;θ) denote ANI-2n given by Eq. (10). If each neural block approximates its modified counterpart with error εi and sub-flows are Lipschitz on the reachable set, then 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\Vert {{{{\mathcal{N}}}}}_{2n}(h;\theta )-{{{\mathcal{U}}}}(h)\right\Vert \,\le \, {\sum}_{i=1}^{n}| {c}_{i}| \,{C}_{i}(h)\,{\varepsilon }_{i},$$\end{document}N2n(h;θ)−U(h)≤∑i=1n∣ci∣Ci(h)εi,where Ci(h) accumulates at most ki Lipschitz factors from the alternating sub-steps (Supplementary Section 5.5). (iii) Global (long-horizon) error analysis. For a final time T = mh and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\beta }_{h}=\min \{{{{\rm{Lip}}}}[{{{{\mathcal{N}}}}}_{2n}(h;\theta )],\,{{{\rm{Lip}}}}[{{{\mathcal{U}}}}(h)]\}$$\end{document}βh=min{Lip[N2n(h;θ)],Lip[U(h)]} on the global reachable set, the m-step error satisfies 14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sup }_{u\in {{{{\mathcal{K}}}}}_{0}}\left\Vert {{{{\mathcal{N}}}}}_{2n}{(h;\theta )}^{m}u-{{{\mathcal{U}}}}{(h)}^{m}u\right\Vert \,\le \,\delta (h,T)\,\mathop{\sum}_{\ell=0}^{m-1}{\beta }_{h}^{\,\ell },\,\delta (h,T):=\left\Vert {{{{\mathcal{N}}}}}_{2n}(h;\theta )-{{{\mathcal{U}}}}(h)\right\Vert,$$\end{document}supu∈K0N2n(h;θ)mu−U(h)mu≤δ(h,T)∑ℓ=0m−1βhℓ,δ(h,T):=N2n(h;θ)−U(h),which yields the exponential-in-time form \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{\delta (h,T)}{Ch}\left({e}^{CT}-1\right)$$\end{document}δ(h,T)CheCT−1 whenever βh≤1 + Ch (Supplementary Section 5.5). Eqs. (13) and (14) are consistent with the observed robustness and clarify how one-step residual errors propagate over long horizons. (iv) Callable prior step maps. If each exact prior sub-flow \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\exp \{\tau {{{\mathcal{A}}}}\}$$\end{document}exp{τA} appearing in the composition, for the substep sizes τ used by ANI, is replaced by a callable map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{P}}}}}_{\tau }$$\end{document}Pτ satisfying \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathop{\sup }_{u\in K}\left\Vert {{{{\mathcal{P}}}}}_{\tau }(u)-\exp \{\tau {{{\mathcal{A}}}}\}u\right\Vert \le {C}_{{{{\mathcal{A}}}}}{\tau }^{q+1}$$\end{document}supu∈KPτ(u)−exp{τA}u≤CAτq+1on the relevant reachable set, then the one-step ANI bound gains an additive prior-discretization contribution obtained by summing over the finitely many prior calls. For fixed composition coefficients and branch counts, this contribution remains of order hq+1. Hence, the same analysis extends from formal flows to executable legacy solvers, reduced numerical pipelines, and frozen learned propagators when such substep error bounds are available (Supplementary Section 5.6). (v) Residual-field realization with classical one-step integrators. If a consistent one-step integrator is used to advance the learned residual field and Φθ approximates the corresponding integrator-exact modified residual field with uniform error εB, then the one-step discrepancy to the target residual flow is controlled by the neural-approximation error. Explicit Runge–Kutta is one instance: if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{{{{\mathcal{B}}}}}}_{h}^{{{{\rm{RK}}}}}$$\end{document}B^hRK denotes an RK-exact modified residual field satisfying \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{h}^{{{{\rm{RK}}}}}({\widehat{{{{\mathcal{B}}}}}}_{h}^{{{{\rm{RK}}}}})=\exp \{h{{{\mathcal{B}}}}\}$$\end{document}ShRK(B^hRK)=exp{hB}, then approximating \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{{{{\mathcal{B}}}}}}_{h}^{{{{\rm{RK}}}}}$$\end{document}B^hRK with uniform error εB yields \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\parallel {u}_{n+1}^{{{{\rm{NN}}}}}-\exp \{h\,{{{\mathcal{B}}}}\}\,{u}_{n}\parallel \le C\,h\,{\varepsilon }_{B}$$\end{document}∥un+1NN−exp{hB}un∥≤ChεB, with C depending only on the chosen tableau, Lipschitz constants, and the step-size upper bound (Supplementary Section 5.5). Other classical one-step schemes can be treated analogously by introducing the corresponding integrator-exact modified residual target, when well defined, and applying the stability bounds of the chosen one-step discretization (Supplementary Section 5.6). (vi) Local Lipschitz stability of ANI on compact reachable sets. Assume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{A}}}},{\widehat{{{{\mathcal{B}}}}}}_{2n}\in {C}^{2}$$\end{document}A,B^2n∈C2 on a neighborhood of the reachable set with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha :=\sup \parallel D{{{\mathcal{A}}}}\parallel$$\end{document}α:=sup∥DA∥ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta :=\sup \parallel D{\widehat{{{{\mathcal{B}}}}}}_{2n}\parallel$$\end{document}β:=sup∥DB^2n∥. Then for sufficiently small h, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{Lip}}}}\,\left[{N}_{2n}(h)\right]\,\le \,1+(\alpha+\beta )\,h+{{{\mathcal{O}}}}({h}^{2}).$$\end{document}LipN2n(h)≤1+(α+β)h+O(h2).In particular, bounded α and β should be interpreted as a local Lipschitz stability condition: over one step, N2n(h) can amplify perturbations by at most \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1+(\alpha+\beta )h+{{{\mathcal{O}}}}({h}^{2})$$\end{document}1+(α+β)h+O(h2). The full proof appears in Supplementary Section 5.3. Post hoc symbolic distillation of learned discrepancy In selected low-dimensional or mechanistically structured settings, the learned ANI correction can be interrogated after training to obtain a simpler hypothesis for the discrepancy. Our procedure is deliberately post hoc: ANI is first trained only for predictive correction, after which we sample state–correction pairs from the learned model, convert the discrete correction into derivative-level targets when needed, and apply sparse regression (SINDy) on a problem-specific library of candidate terms. Not every discrepancy is expected to admit a compact symbolic form; when such a form is plausible, this workflow provides a route to refinement. In this work, the procedure is demonstrated on the Lorenz–Stenflo, FitzHugh–Nagumo, and battery-aging examples; the resulting expressions are reported in Supplementary Section 9. In more complex settings, the learned neural correction itself remains a valid effective scientific object even without a closed-form distillation. Supplementary information Supplementary Information Transparent Peer Review file Supplementary Information Transparent Peer Review file