Technical Summary: Physics-Encoded Inverse Modeling for Arctic Snow Depth Prediction

1. Problem Statement

The paper addresses the challenge of estimating Arctic snow depth, a critical but latent variable in the Arctic Ocean system. This task is formulated as a time-varying inverse problem characterized by:

• Data Scarcity: Direct observations of snow depth are unavailable in standard reanalysis datasets (e.g., ERA5). Available data (snow density, sea ice concentration, albedo) are sparse, indirect, and often spatially averaged.
• Ill-Posed Nature: The mapping from sparse observations to hidden physical parameters is not bijective (one-to-one). Multiple latent physical configurations can produce similar observational averages, making standard inversion difficult.
• Limitations of Existing Methods:
  • Physics-Informed Neural Networks (PINNs): Typically rely on complex Partial Differential Equations (PDEs) and residual computations, which are computationally expensive and often unnecessary for systems governed by linear structural dependencies (like hydrostatic equilibrium).
  • Standard Inverse Modeling: Often assumes perfect invertibility or requires extensive hyperparameter tuning in self-supervised learning (SSL) settings, failing to handle noisy, incomplete real-world data robustly.

2. Methodology: PhysE-Inv Framework

The authors propose Physics-Encoded Inverse Modeling (PhysE-Inv), a unified framework combining deep sequential learning with physics-encoded inference. The architecture is end-to-end differentiable and consists of four main components:

A. Data and Proxy Target

• Input: Time series of normalized snow density ($X$) and other ERA5 variables (sea ice concentration, albedo).
• Target: Since direct snow depth ($h_s$) is missing, a proxy target is generated using the hydrostatic balance equation:
  $$\rho_i h_i + \rho_s h_s = \rho_w (h_i - h_f)$$
  This relationship links ice thickness, snow depth, and freeboard, allowing the construction of a target variable ($h_{proxy}$) from available density and concentration data.

B. Model Architecture

1. Encoder-Decoder with Attention:
   • LSTM Encoder: Captures temporal dependencies in the input sequence.
   • Multi-head Self-Attention: Refines encoder outputs to capture non-local temporal interactions, crucial for sparse observations.
   • LSTM Decoder: Integrates refined information into a global latent representation ($z_n$).
2. Parameter Estimation Module (Inverse Modeling Block):
   • Instead of direct regression, the model estimates three learnable physical parameters ($w, b, c$) via a Multi-Layer Perceptron (MLP) from the latent state $z_n$.
   • Differentiable Constraints: Parameters are transformed to ensure physical bounds (e.g., $w \in [-1, 1]$ for density-depth coupling, $b > 0$ for confidence scaling, $c$ for bias correction).
   • Physics-Encoded Reconstruction: The final prediction is not just a neural output but a structured combination:
     $$\hat{y}_{est} = w \bar{\rho}_s + b \hat{y}_{pred} + c$$
     Where $\bar{\rho}_s$ is the mean input density and $\hat{y}_{pred}$ is the direct neural prediction. This enforces the linear structural dependency of hydrostatic balance.
3. Contrastive Learning Regularization:
   • To handle noise and data scarcity, the framework employs NT-Xent contrastive loss.
   • It maps clean sequences and their noisy augmentations to identical points in the latent space, enforcing noise-invariant representations. This is integrated directly into training rather than used as a pretraining step.

C. Loss Function

The total loss combines three terms:
$$L_{total} = L_{depth} + L_{physics} + \lambda_{contrast} L_{contrast}$$

• $L_{depth}$: MSE between direct neural predictions and ground truth.
• $L_{physics}$: MSE between the physics-structured reconstruction and ground truth.
• $L_{contrast}$: Regularization for latent robustness.

3. Key Contributions

1. Unified Framework: Introduces PhysE-Inv, bridging sequential deep learning (LSTM/Attention) with physics-encoded inference specifically for linear structural dependencies (hydrostatic balance), avoiding the computational overhead of PDE-based PINNs.
2. Surjective Inverse Mapping: Proposes a formulation that accommodates many-to-one latent states (surjective mapping), which is more realistic for spatially averaged Arctic observations than bijective assumptions.
3. Integrated Contrastive Learning: Embeds contrastive regularization directly into the training loop to learn noise-invariant features, addressing the challenge of data scarcity without explicit labels.
4. Differentiable Parameter Estimation: Enables robust recovery of physical parameters ($w, b, c$) while preserving gradient flow, ensuring the estimated hidden parameters remain physically meaningful.

4. Experimental Results

The framework was evaluated on ERA5 data (2019–2023) over the central Arctic Ocean against four baselines: LSTM, BiLSTM, Neural ODE, and ResNet-50. All baselines were augmented with the same physics-encoded head and contrastive strategy for fair comparison.

• Performance Gains:
  • PhysE-Inv achieved the lowest error rates, with an MSE of 0.3568 and RMSE of 0.5973 when using parameter estimation.
  • This represents a 24.7% average MSE reduction across all baselines and a 17.3% improvement over the strongest baseline (ResNet-50) under parameter estimation settings.
  • The inclusion of the parameter estimation module reduced MSE by 9.5% for PhysE-Inv specifically.
• Impact of Contrastive Learning (CL):
  • CL provided consistent improvements across varying data sparsity regimes (80%, 60%, and 50% of data).
  • In the most data-scarce regime (50%), CL still reduced MSE by 8.5%.
  • In the 80% regime, CL reduced MSE by 10.2%, demonstrating its ability to leverage additional data effectively.
• Distributional Analysis:
  • PhysE-Inv showed the closest alignment with the ground truth distribution of snow depth anomalies, with a median near zero and minimal outliers.
  • Baselines exhibited issues such as excessive variability (BiLSTM), underestimation of variance (ResNet-50), or upward bias (Neural ODE).
• Temporal Robustness:
  • Position-wise RMSE analysis showed PhysE-Inv outperforming all baselines at every time step, with the performance gap widening as temporal context accumulated.

5. Significance and Claims

The paper claims that PhysE-Inv establishes a generalizable inverse modeling paradigm for data-scarce domains where physics-informed guidance can be incorporated into sparse observations.

• Addressing the Mismatch: It resolves the mismatch between heavy PDE-based frameworks (PINNs) and systems governed by simpler linear relationships, offering a "lighter, physically grounded alternative."
• Robustness to Sparsity: By combining contrastive learning with physics constraints, the model achieves robust feature extraction even when labeled data is extremely limited.
• Physical Interpretability: Unlike black-box deep learning, the framework explicitly estimates parameters ($w, b, c$) that have clear physical interpretations (e.g., density-depth coupling, confidence scaling, bias correction), ensuring the hidden parameters are physically meaningful in the context of Arctic sea ice.

The authors conclude that this approach effectively solves ill-posed inverse problems in environmental science where observations are sparse, paving the way for future extensions involving Bayesian uncertainty quantification and validation against in-situ measurements.