Authors: Chawit Kaewnuratchadasorn, Jiaji Wang, Xin Nie
Organizations: Department of Civil Engineering, The University of Hong Kong, Pok Fu Lam, Hong Kong, China. · Department of Civil Engineering, Tsinghua University, Haidian, Beijing, China.
High-fidelity simulation of dynamical systems is central to scientific discovery and engineering design, yet remains computationally demanding. Here we introduce the Implicit Neural Operator Solver (INOS), which learns the implicit solve and encodes a time integrator. INOS pretrains a neural operator from static equilibrium solutions and reuses it within Newmark-β integration. At each time increment, the integrator forms an effective equilibrium problem, the pretrained neural operator infers the next-step solution field, and the Newmark relations advance the dynamic state. Numerical examples in linear and nonlinear elastodynamics demonstrate accurate full-field predictions without dynamic trajectory supervision. Comparative results show that INOS, evaluated zero-shot, achieves higher predictive accuracy than neural solver baselines trained on dynamic simulations. In addition, INOS delivers over 4,600× speedup in nonlinear simulations over sequential classical finite-element solvers. INOS natively accommodates changes in dynamic conditions and enables reuse of the pretrained operator across spatial and temporal discretizations. Further evaluation demonstrates a stable million-step rollout. These findings establish INOS as a reusable neural implicit solver for accurate, efficient, and robust dynamic modeling.
Figures & tables
Figure 1: Implicit Neural Operator Solver (INOS): learning static equilibrium for zero-shot implicit dynamics. a , After spatial discretization, Newmark integration reformulates each dynamic step as static equilibrium with added stiffness Kadddyn=a0M+a1C , where M and C represent mass and damping, and a0,a1 are Newmark coefficients. b , A neural operator Gθ learns the equilibrium map from static problems with added nodal springs, without dynamic training data. c , In autoregressive rollout, the predicted state sn and external load determine the effective force (1), the operator predicts displacement (2), and the Newmark update advances velocity and acceleration (3). Only the implicit solve is learned through static equilibrium. The time integration follows the prescribed scheme.
Figure 2: INOS rollouts on linear elastodynamics. a , 2D panel. b , 2D perforated panel. c , 2D multistory shear wall. d , 3D solid. Displacement magnitudes ∣u∣ , histories at the displayed nodes and cumulative errors are compared with finite-element references (Ref.) using the same spatial discretization, time increment, and Newmark parameters. DOFs, degrees of freedom; Δt , time increment.
Figure 3: INOS rollouts on nonlinear elastodynamics. a , 1D Neo-Hookean bar: axial displacement u(x,t) . b , 2D Neo-Hookean panel: displacement magnitude ∣u∣ at reported snapshots. Fields, nodal histories, and cumulative errors compare INOS with Newton FE references (Ref.) using the same spatial discretization and Newmark parameters. λ , reference principal stretches over the rollout; L , bar length; Δt , time increment; DOFs, degrees of freedom.
Figure 4: Comparison with neural solver baselines trained on dynamic data. a , 1D dynamics: axial displacement u(x,t) . b , 2D dynamics: displacement magnitude ∣u∣ ; vertical dashed lines mark snapshot times. INOS uses static pretraining and is evaluated zero-shot on dynamic problems, whereas FNO, LSTM/ConvLSTM and Transolver are trained on dynamic trajectories. Tests are IID (independent and identically distributed) or OOD (out of distribution) in force amplitude (F) or rollout horizon (H), relative to the baselines’ training distribution. Ref., FE reference; DOFs, degrees of freedom.
Figure 5: Computational efficiency of INOS. a , Wall-clock rollout times for the 2D linear benchmarks in Fig. 2 . b , Wall-clock times for 50-step rollouts of the 2D Neo-Hookean panel. c , Total wall-clock time for N nonlinear simulations, run sequentially by the FE solvers and as a single GPU batch by INOS. Annotations at N=64 show INOS speedups over the FE solvers.
Figure 6: Generalization and rollout stability. a , 1D dynamical responses under applied loading, free release from an initial condition, and loading followed by release, at the indicated time increments Δt . Space-time maps show INOS predictions at Δt=1 ms; nodal histories compare predictions with FE references (Ref.) at the marked node. b , A 106 -step rollout at Δt=1 ms: nodal displacement ranges within 5-s windows, the first and last 1,000 steps, and cumulative full-field error. All predictions were obtained from the same pretrained INOS solver. DOFs, degrees of freedom.
Rel. ℓ2 [%] ↓
Model
Evolution
Params
Sta
IID
OOD(F)
OOD(H)
Dynamic trajectory learning: Fext1:T→U1:T
FNO
learned (x,t) map
2.41M
–
3.57
6.34
35.96 ∗
LSTM
learned recurrence
3.88M
–
2.90
13.58
10.24
Transolver
learned autoregression
1.13M
–
1.92
14.22
9.74
Static equilibrium learning and implicit time integrator: Feffn+1→un+1
Table 1: 1D dynamic benchmark. Relative ℓ2 error (%) of trajectory-trained baselines and static-trained INOS on held-out static equilibrium samples (Sta) and in trajectory-level rollouts under IID, OOD(F), and OOD(H) conditions; lower is better. Sta is not applicable to the trajectory-trained baseline.
Rel. ℓ2 [%] ↓
Model
Evolution
Params
Sta
IID
OOD(F)
OOD(H)
Dynamic trajectory learning: Fext1:T→U1:T
FNO
learned (x,y,t) map
18.89M
–
7.96
10.34
9.42 ∗
ConvLSTM
learned recurrence
1.05M
–
3.33
4.87
8.34
Transolver
learned autoregression
1.42M
–
6.14
10.06
3.06
Static equilibrium learning and implicit time integrator: Feffn+1→un+1
Table 2: 2D dynamic benchmark. Relative ℓ2 error (%) of trajectory-trained baselines and static-trained INOS on held-out static equilibrium samples (Sta) and in trajectory-level rollouts under IID, OOD(F), and OOD(H) conditions; lower is better. Sta is not applicable to the trajectory-trained baseline.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Title
Key content and purpose
5
5 Theoretical formulation
5.1 : 5.1 Elastodynamic theory and INOS formulation 5.2 : 5.2 Implementation details of INOS rollout 5.3 : 5.3 Error decomposition in INOS 5.4 : 5.4 Rollout time complexity 5.5 : 5.5 Differentiability of the rollout map
6
6 Problem specifications, data and reference solutions
Table 3: Overview of the Supplementary Information.
Backbone
One evaluation Cnet(N)
Rollout Croll(N,T)
FNO Li et al. (2021a)
O(L[cNlogN+c2N])
O(TNlogN)
Transolver Wu et al. (2024)
O(L[cNM+cM2+c2N])
O(TN)
Appendix
Table 4: Backbone evaluation and rollout complexity.
1D truss
2D panel
2D perforated panel
2D shear wall
3D solid
Structure
Size [m]
24
1×1
1×1
24×36
4
Section
A=0.006 m 2
t=0.01 m
t=0.01 m
t=0.25 m
0.25×0.25 m
Geometry
–
–
hole
12 floors
–
Elas. modulus E [GPa]
200
210
210
30
200
Poisson’s ratio ν
–
0.30
0.30
0.20
0.30
Appendix
Table 5: Configurations of the linear elastodynamic benchmarks. Operator inputs are the force and added-stiffness fields (three components each for the 3D solid). Nt includes the initial state. GRF, Gaussian random field; DOFs, degrees of freedom; NR, not reported.
1D Neo-Hookean bar
2D Neo-Hookean panel
Structure
Size [m]
0.5
1×1
Section
A=10−4 m 2
t=0.01 m
Elastic constants
μ=0.6 MPa
E=1.5 MPa, ν=0.45
Density ρ [kg m -3 ]
1,030
1,100
DOFs
101
2,178
Appendix
Table 6: Configurations of the nonlinear (Neo-Hookean) elastodynamic benchmarks. μ , shear modulus; kˉadd , added stiffness per unit reference length; T1 , first natural period; FC, fully connected layer; other entries as in Supplementary Table 5 .
Figure 7: Spatial discretization checks for linear benchmarks. Relative differences from the finest reference mesh in each case: ( a ) displacement trajectories at a fixed time step; ( b-d ) displacement and strain energy from auxiliary linear-static tests. Stars highlight selected mesh resolutions.
Table 7: Neural model configurations. All models are trained using the Adam optimizer.
Figure 8: Static-pretraining coverage and dynamic evaluation regimes. ( a , d ) Representative static and dynamic effective-equilibrium data for the 1D and 2D benchmarks, respectively. ( b , e ) Empirical force and stiffness input distributions for static pretraining, IID dynamics and OOD(F), with dynamic forces expressed as effective-equilibrium inputs. ( c , f ) Representative time histories for IID and OOD(H). (labeled OOD(T) in the legends).
Figure 9: Coverage-controlled design of the static-to-dynamic transfer experiment. (a) Distribution of static stiffness inputs and Newmark-induced effective-stiffness ratios across the tested time steps. (b) Dynamic cases (applied forces), shown through the nodal components fx and fy at a representative node on the 32×32 panel under Δt=0.02 s and Δt=0.0015 s.
Δt [s]
N
ρinput
erollT
eroll25
0.0200
25
0.031
0.01397
0.01397
0.0100
50
0.081
0.01416
0.01332
0.0080
63
0.114
0.01414
0.01320
0.0050
100
0.242
0.01433
0.01273
0.0030
167
0.581
0.01473
0.01170
0.0025
200
0.804
0.01520
0.01155
Appendix
Table 8: Variable- Δt rollout accuracy of the static-pretrained INOS solver.
Figure 10: Variable- Δt rollout accuracy of the static-pretrained INOS solver. (a) rollout errors versus coverage ratio ρinput . (b) Representative displacement responses.
Figure 11: Spatial resolution transfer in 1D linear elastodynamics. Columns show the native, 2× and 4× resolutions. Rows 1-3 show displacement fields from the FE reference and INOS with FNO and Transolver backbones. Rows 4-5 show midpoint displacement histories and cumulative errors, respectively. Annotations report rollout relative ℓ2 errors.
Neural PDE solvers use trainable models to approximate solutions or solution operators for partial differential equations, often with physical constraints.
Solving dynamical systems in complex and diverse geometries remains challenging for both numerical solvers and learning-based surrogate models. We propose Neural Operator Splitting, a framework for learning fluid dynamics while preserving the physical structure of fractional-step numerical schemes. Instead of learning the full-step evolution monolithically, we decompose each time step into physically interpretable steps, and learn each operator separately using intermediate states from the underlying numerical solver. Guided by the linearity properties of the fractional steps, we introduce a conditional-linear Fourier neural operator (CL-FNO) whose linear state mappings are modulated by nonlinear conditioning on inputs such as boundary geometry. We further introduce two complementary strategies, symmetry sampling and corrective cycling, to mitigate error accumulation during autoregressive rollout. Experiments on 2D and 3D fluid flows with varying geometries and initial conditions show that our full model reduces mean relative L2 rollout error by 68% in 2D and 40% in 3D compared with a single full-step neural operator. At comparable error levels, neural-operator inference is approximately 103 times faster per step than the Monte Carlo solver.
Neural PDE solvers can achieve low prediction errors, but do they reproduce the dynamics of the systems they model? Prediction scores alone offer an incomplete answer: they measure agreement with reference solutions but provide limited insight into how errors accumulate, nearby states diverge, or extreme events arise. We propose an evaluation framework that directly examines these behaviors in deterministic and stochastic neural solvers. By evolving ensembles of nearby initial states and comparing them with direct numerical simulation, we assess three complementary aspects of learned dynamics: error formation, ensemble geometry, and extreme events. Experiments on two-dimensional Kolmogorov flow reveal limitations that conventional scores can obscure. Smaller trajectory errors can reflect weaker error amplification despite less accurate local updates. Models can match an ensemble's overall spread and effective dimension while failing to capture the spatial directions where nearby states diverge. Similarly, matching overall event frequencies can conceal failures to predict persistent extreme events. These findings show that improved prediction accuracy does not necessarily imply greater dynamical fidelity. Our framework makes this distinction measurable, providing concrete criteria for evaluating whether advances in neural PDE solvers better capture the underlying dynamics.
Haonan Li, Yue Song, Bin Yang +1
Center for Combustion Energy, Tsinghua University · College of AI, Tsinghua University · Department of Mechanical Engineering, University College London
We propose neural network operator inference (NN-OpInf): a structure-preserving, composable, and minimally restrictive operator inference framework for the non-intrusive reduced-order modeling of dynamical systems. The approach learns latent dynamics from snapshot data, enforcing local operator structure such as skew-symmetry, (semi-)positive definiteness, and gradient preservation, while also reflecting complex dynamics by supporting additive compositions of heterogeneous operators. We present practical training strategies and analyze computational costs relative to linear and quadratic polynomial OpInf (P-OpInf). Numerical experiments across several nonlinear and parametric problems demonstrate improved accuracy, stability, and robustness over P-OpInf and prior NN-ROM formulations, particularly when the dynamics are not well represented by polynomial models. These results suggest that NN-OpInf can serve as an effective drop-in replacement for P-OpInf when the dynamics to be modeled contain non-polynomial nonlinearities, offering potential gains in accuracy and out-of-distribution performance at the expense of higher training computational costs and a more difficult, non-convex learning problem.
Eric Parish, Anthony Gruber, Patrick Blonigan +1
Sandia National Laboratories, Livermore, CA, USA · Sandia National Laboratories, Albuquerque, NM, USA