Technical Summary: Recovering Physical Dynamics from Discrete Observations via Intrinsic Differential Consistency

1. Problem Statement

Learning continuous-time physical dynamics from discrete, sparse, and noisy observations remains a fundamental challenge in scientific machine learning (SciML). Existing paradigms face a dichotomy:

• Discrete-time state regression (e.g., auto-regressive models) is bound by fixed time steps, leading to temporal rigidity and poor generalization across varying horizons.
• Neural ODEs learn continuous dynamics but often produce numerically plausible yet physically inconsistent flows due to a lack of structural supervision between observations. They rely on local truncation errors for adaptive stepping, which assumes the governing equations are exact—a condition often violated by neural surrogates.
• Physics-Informed Neural Networks (PINNs) enforce physical compliance via explicit equations but require knowledge of the governing PDEs, which is frequently unavailable.

The core limitation across these approaches is a reliance on local constraints (pointwise regression, derivative approximations, or equation residuals). As the observation interval grows, the fidelity of these local supervision signals degrades, leading to compounding errors and representational collapse during long-horizon extrapolation.

2. Methodology: Consistent Vector Flow (CVF)

The authors propose Consistent Vector Flow (CVF), a framework that treats dynamic evolution not as a sequence of discrete states but as a continuous flow satisfying Time-Translation Symmetry (TTS). The method is built on three pillars:

A. Hypothesis Space and Symmetry Rupture

The authors define a hypothesis space $\mathcal{H}$ of flows that satisfy the semi-group property under time translation. For an autonomous system, the evolution operator $\Phi_t$ must satisfy $\Phi_{t_1+t_2}(s) = \Phi_{t_2}(\Phi_{t_1}(s))$.

• Symmetry Rupture ($R_k$): The deviation from this property is quantified as the residual between a multi-step composed flow and a direct flow prediction. Specifically, the paper focuses on the triangular loop ($k=3$), where the sum of two sequential steps should equal the direct step.
• Training Objective: The model minimizes a loss function combining the secant velocity matching error and the Symmetry Rupture ($R_3$). This forces the learned representations to compose consistently across different temporal scales.

B. Temporal Expansion of the Velocity Field

To resolve the representational inconsistency between discrete observations and continuous tangent velocity fields, CVF learns a time-conditioned secant velocity field $\psi_\theta(s_t, \Delta t)$ rather than a tangent field $v_\theta(s_t)$.

• The model is trained on the secant velocity: $\psi_\theta(s_t, \Delta t) = \frac{\Phi_{\Delta t}(s_t) - s_t}{\Delta t}$.
• This approach eliminates systematic bias in tangent approximation and allows the model to learn building blocks of dynamics at multiple temporal scales. The true dynamics ($\Delta t \to 0$) become a sub-problem of learning a consistent cross-temporal bundle representation.

C. Cascaded Normalization

To stabilize training and inference, the authors introduce a cascaded normalization scheme. Unlike standard independent normalization of states and velocities, this method applies a pushforward transformation to the velocity field based on the state normalization. This preserves the intrinsic geometric relationship between the base manifold and its tangent space, preventing gradient mismatches in stiff systems.

D. Adaptive Consistency Solver (Inference)

The paper repurposes the Symmetry Rupture as an intrinsic consistency-aware oracle for adaptive step-size control during inference, replacing the reliance on local truncation error.

• Mechanism: The solver dynamically searches for the largest macroscopic step size $\Delta t$ such that the normalized Symmetry Rupture remains below a tolerance.
• Principles:
  1. Minimum Step Size Constraint: Steps cannot be smaller than the training data resolution ($\delta_{min}$).
  2. Dynamic Relative Tolerance: Tolerance adapts to the current operational scale, shifting from spatial error control to temporal error control.
• This allows the solver to allocate compute based on local geometric complexity, taking large steps in consistent regions and subdividing only where representational consistency breaks down.

3. Key Contributions

1. Consistency-Based Hypothesis Space: Grounding the learning of physical dynamics on a TTS-guided hypothesis space, enabling the recovery of continuous dynamics without explicit equation supervision.
2. Temporal Expansion for Unbiased Learning: Leveraging a time-conditioned secant velocity field to eliminate systematic bias in tangent approximation and mitigate conflicting learning objectives between true dynamics and flat surrogate extrapolation.
3. Unified Training and Inference Framework: Utilizing the same consistency metric (Symmetry Rupture) as both a training regularizer to constrain the hypothesis space and as a step-size scheduling oracle during inference.
4. Adaptive Consistency Solver: An adaptive solver that uses the model's own consistency estimate as a local error indicator, dynamically adjusting integration step sizes to improve accuracy and efficiency.

4. Experimental Results

The method was evaluated on public PDE benchmarks (2D Diffusion-Reaction, 2D Shallow Water, 2D CFD Turbulence, and 2D Wave Equation) against state-of-the-art baselines including Neural ODEs (N-ODE), Flow Matching (CFM), and derivative matching methods.

• Time-Informed Inference: On the Diffusion-Reaction benchmark, CVF reduced rollout RMSE by 87% compared to baselines while using 5x fewer function evaluations (NFE). CVF achieved a rejection-free, single-step rollout behavior (NFE $\le$ 3) across all datasets, whereas baselines relied heavily on high-order solvers (NFE $\ge$ 20).
• Direct Auto-regressive Inference: In the more demanding setting where models predict distant future frames without intermediate temporal cues, CVF maintained the lowest rollout RMSE on two of three benchmarks.
  • Baselines like N-ODE either diverged or required up to an order of magnitude more function evaluations (e.g., NFE jumping from 20.1 to 115.9 on Shallow Water) to remain stable, indicating a reliance on brute-force micro-stepping to compensate for structural flaws.
  • CVF successfully allocated compute based on local geometric complexity, maintaining stability with significantly lower NFE.
• Ablation Studies:
  • Downsampling: Uniform downsampling during training was shown to be critical. Random downsampling led to "overconfident" extrapolation where the model collapsed to a linearized vector field, failing to detect representational inconsistency.
  • Normalization: Cascaded normalization was proven essential for stabilizing training in stiff systems.
  • Symmetry Type: Enforcing semi-group properties (unidirectional) was superior for dissipative systems, while bidirectional constraints were comparable for conservative systems.

5. Significance and Claims

The paper claims that a sophisticated solver cannot rescue a structurally flawed flow. Existing methods often learn biased representations that deviate from the true physical phase space; applying high-order solvers merely integrates these erroneous fields more precisely, leading to systematic errors.

By constraining the hypothesis space to TTS-compatible flows, CVF ensures that the learned vector field is geometrically consistent with valid physical flows before the solver steps in. This approach:

• Decouples computational cost from error accumulation: By learning a consistent geometric prior, the model can take large macroscopic steps without compounding errors, breaking the traditional trade-off between efficiency and accuracy.
• Provides a new inductive bias: It highlights the importance of embedding physical domain knowledge (specifically differential consistency and semi-group properties) directly into the loss landscape, rather than relying solely on architecture or optimization design.
• Offers a robust alternative to equation-based methods: It recovers continuous dynamics from discrete data without requiring explicit knowledge of the governing PDEs, addressing a key limitation of PINNs.

The authors conclude that while their framework focuses on autonomous systems and shows marginal training overhead, it yields significant reductions in long-horizon prediction error and function evaluations, establishing a new paradigm for learning consistent physical dynamics from discrete observations.