🤖 machine learning

Knowledge-Guided Time-Varying Causal Inference for Arctic Sea Ice Dynamics

This paper proposes the Knowledge-Guided Causal Model Variational Autoencoder (KGCM-VAE), a framework that integrates physical constraints and Maximum Mean Discrepancy to accurately quantify the causal effect of sea surface height on sea ice thickness by addressing time-varying confounding and outperforming existing deep learning baselines.

Original authors: Akila Sampath, Vandana Janeja, Jianwu Wang

Published 2026-04-22

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Akila Sampath, Vandana Janeja, Jianwu Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Why Do We Need This?

Imagine you are trying to figure out why a specific plant in your garden is wilting. You notice that every time the wind blows hard (Sea Surface Height), the plant droops (Sea Ice Thickness). But is the wind actually killing the plant? Or is it just that the wind happens to blow when the soil is also dry (a hidden factor)?

In the real world, especially in the Arctic, things are messy. The ocean, the ice, the wind, and the temperature all change every second. Old computer models are like rigid rulebooks; they assume the wind always affects the ice in the exact same way. But in reality, the relationship is fluid and changes depending on the season or the current speed of the water.

The authors of this paper built a new, smarter AI called KGCM-VAE. Think of it as a "Super-Detective" that doesn't just look at patterns, but understands the physics of the situation to figure out what is truly causing what.

The Problem: The "Chicken or Egg" of the Arctic

Scientists know that Sea Surface Height (SSH) (how high the water is) and Sea Ice Thickness are connected. But it's a tangled web:

Confounding: A hidden force (like deep ocean currents or atmospheric pressure) might be pushing the water up and melting the ice at the same time. If you don't account for this, you might blame the water height for something the wind actually did.
Time-Varying: The effect isn't constant. A small push on the ice today might do nothing, but if the ocean is already moving fast, that same push could trigger a massive collapse. It's like pushing a swing; if it's already moving, a tiny nudge sends it flying. If it's still, the nudge does nothing.
Continuous: The "push" (treatment) isn't just "on" or "off." It's a dial that can be turned up or down infinitely.

The Solution: The "Physics-Savvy" AI

The authors created KGCM-VAE. Let's break down its superpowers using a metaphor:

1. The "Physics Guidebook" (Knowledge-Guided)

Most AI learns by guessing based on data alone. This AI, however, has a Physics Guidebook in its pocket.

The Analogy: Imagine a chef learning to cook. A normal AI tastes a dish and guesses the recipe. The KGCM-VAE is a chef who also knows the laws of chemistry. It knows that if you heat water, it boils. It uses these known laws (like how water density affects ice floating) to guide its guesses.
In the paper: They use real physics equations (like how ice floats based on water density) to create "treatments." Instead of randomly changing the water height, the AI simulates changes that could physically happen in the real ocean.

2. The "Time Machine" (Time-Varying Causal Inference)

The AI looks at the past 30 days of data to predict the future.

The Analogy: Think of a car driver. If you want to know if a car will crash, you don't just look at where it is now. You look at how fast it was going 10 seconds ago, 20 seconds ago, and how the road was curving.
In the paper: The model remembers the history of the ocean currents. It understands that a change in water height today might not affect the ice until three days later. It calculates the "Individual Treatment Effect" (ITE), which is basically asking: "If we had nudged the water height slightly differently yesterday, how much thicker or thinner would the ice be today?"

3. The "Fairness Filter" (MMD Regularization)

This is the most technical part, but here's the simple version.

The Problem: In real life, we can't run experiments where we freeze the ocean in one place and melt it in another to compare. We only have one history. This creates "bias." It's like trying to see if a new medicine works, but the people who took the medicine were all young and healthy, while the people who didn't take it were old and sick. You can't tell if the medicine worked or if the age difference caused the result.
The Solution (MMD): The AI uses a mathematical trick called Maximum Mean Discrepancy (MMD).
The Analogy: Imagine you have two groups of students: Group A studied with a new method, and Group B studied with the old method. But Group A also had better teachers. To compare fairly, the AI uses MMD to "rearrange" the students in its mind so that Group A and Group B look statistically identical in every way except for the study method. This forces the AI to focus only on the effect of the study method, ignoring the "better teachers" (confounders).

What Did They Find? (The Results)

The team tested their AI in two ways:

The Simulation Test (Synthetic Data): They created a fake, perfect world where they knew the exact answer.
- Result: KGCM-VAE was the best detective. It predicted the ice thickness changes more accurately than any other existing AI model. It proved that adding the "Fairness Filter" (MMD) made the predictions much sharper.
The Real-World Test: They applied it to real Arctic data.
- The "Tipping Point" Discovery: The AI found something fascinating.
  - Scenario A (Velocity): When they simulated changing the speed of ocean currents, the ice thickness barely changed. Why? Because the currents weren't strong enough to cross the "tipping point." It's like pushing a heavy boulder; if you don't push hard enough, it doesn't move.
  - Scenario B (Water Height): When they simulated changing the water height, the ice thickness started to wiggle up and down wildly. Why? Because the water height changes were strong enough to push the system over the edge, causing the ice to melt and then recover in a cycle.
- Conclusion: The AI correctly identified that small changes sometimes do nothing, but once you cross a specific threshold, the system reacts violently. This is a "non-linear" response that older, simpler models missed.

The Takeaway

This paper introduces a new way to use AI for climate science. Instead of just letting the AI guess patterns from messy data, they gave it a rulebook of physics and a fairness filter.

In short: They built a smarter AI that understands that the Arctic Ocean is a complex, changing system. It can now tell us not just what is happening to the ice, but why it's happening, and what might happen if we nudged the ocean in a different direction. This helps scientists predict sea-level rise and climate change with much greater confidence.

1. Problem Statement

The paper addresses the critical challenge of quantifying the causal relationship between Sea Surface Height (SSH) and Arctic Sea Ice Thickness (SIT). While SSH is a key indicator of sea-level rise and ocean circulation, its non-direct influence on sea ice thickness (mediated by ocean circulation and freshwater distribution) is difficult to quantify.

Key Challenges Identified:

Time-Varying Confounding: Traditional causal inference methods struggle with climate data where confounders (e.g., atmospheric drivers, geostrophic flow) change over time and influence both the treatment (SSH) and the outcome (SIT).
Continuous & Lagged Treatments: Most existing temporal causal models assume discrete treatments. However, SSH forcing is continuous and induces delayed effects on sea ice, which standard models fail to capture effectively.
Lack of Physical Constraints: Purely data-driven deep learning models often learn spurious correlations because they lack the physical priors necessary to constrain the causal search space to mechanically plausible relationships.
Unmeasured Confounders: Hidden factors (like unobserved geostrophic flow) create spurious correlations that obscure true causal pathways.

2. Methodology: KGCM-VAE

The authors propose the Knowledge-Guided Causal Model Variational Autoencoder (KGCM-VAE), a specialized architecture designed to estimate Individual Treatment Effects (ITE) in sequential, time-varying systems.

A. Physical Knowledge Integration

The model is grounded in established Arctic physics:

Hydrostatic Equilibrium: The relationship between sea ice thickness ( $h_i$ ), sea surface height ( $h_{SSH}$ ), and densities of water, ice, and snow is used to ensure physical consistency.
Geostrophic Balance: The model leverages the relationship between SSH gradients and surface velocity ( $u_g, v_g$ ) to identify confounders.
Causal Graph: A Directed Acyclic Graph (DAG) is constructed where Total Velocity acts as an observed proxy for the unmeasured confounder (geostrophic flow), allowing for the blocking of backdoor paths between SSH and SIT.

B. Knowledge-Guided Treatment Generation

Instead of using raw data as treatment, the model generates physically grounded, time-varying continuous treatments:

Mechanism: A moving window filter smooths the SSH signal. This signal is then dynamically amplified based on ocean circulation intensity (total surface velocity).
Non-linearity: A sigmoid function is employed to model a "tipping-point" dynamic. If surface velocity is below a critical threshold ( $v_0$ ), the amplification is low; above the threshold, the forcing on sea ice is significantly amplified. This mimics real-world regime shifts where small perturbations can trigger disproportionate responses.

C. Model Architecture

The framework extends a standard Variational Autoencoder (VAE) with the following components:

Encoder ( $q(Z|X, T)$ ):
- Inputs: Historical covariates ( $X_t$ : velocities, SSH) and treatment ( $T_t$ ).
- Mechanism: Uses a Bidirectional GRU (Bi-GRU) to capture forward and backward temporal dependencies.
- Output: A latent representation $Z_t$ that is "treatment-aware."
Decoder ( $p(Y|Z, T)$ ):
- Inputs: Latent state $Z_t$ and the treatment value $T_t$ (projected into an embedding).
- Mechanism: A Multi-Layer Perceptron (MLP) predicts the potential outcome (SIT).
- Counterfactual Inference: By swapping the treatment input ( $T_0$ vs. $T_1$ ) while keeping the latent state $Z_t$ constant, the model predicts factual and counterfactual outcomes to calculate the ITE.
Deconfounding via MMD:
- To address unmeasured confounding, the model applies Maximum Mean Discrepancy (MMD) regularization in the latent space.
- It forces the latent distributions of the "treated" and "control" groups to be statistically indistinguishable, effectively balancing the confounders without explicit propensity score estimation.

D. Training Objective

The model is trained end-to-end by minimizing a composite loss function:
$\mathcal{L}_{total} = \beta_{rec}\mathcal{L}_{MSE} + \beta_{KL}\mathcal{L}_{KL} + \beta_{MMD}\mathcal{L}_{MMD}$
Where $\mathcal{L}_{MSE}$ is reconstruction error, $\mathcal{L}_{KL}$ is the KL divergence (regularizing the latent space), and $\mathcal{L}_{MMD}$ enforces treatment balance.

3. Key Contributions

Novel Architecture: Introduction of KGCM-VAE, the first framework to integrate physics-guided treatment generation with VAE-based causal representation learning for time-varying continuous treatments.
MMD for Temporal Deconfounding: Demonstration that MMD regularization in the latent space effectively mitigates observed confounding bias in sequential settings, outperforming traditional propensity score methods which struggle with continuous, high-dimensional time series.
Physical Consistency: The model successfully captures non-linear, threshold-dependent dynamics (tipping points) observed in ocean-ice interactions, which linear models miss.
Robust Estimation: The framework provides unbiased estimates of causal effects even in the presence of unmeasured seasonal confounders.

4. Experimental Results

The model was evaluated using both synthetic data (with known ground truth) and real-world Arctic data.

Synthetic Data Performance:
- Metric: Precision in Estimation of Heterogeneous Effects (PEHE).
- Comparison: KGCM-VAE outperformed state-of-the-art baselines (CRN, Causal-RNN, G-Net) across all treatment lags (3, 6, and 9 days).
- Ablation: Removing the MMD component resulted in a consistent increase in PEHE, confirming that MMD is crucial for reducing confounding bias.
Real-World Case Study:
- Velocity Perturbation: When velocity was perturbed below the critical threshold, the model correctly predicted negligible changes in sea ice thickness (consistent with tipping-point theory).
- SSH Perturbation: When SSH was perturbed, the model captured complex, oscillatory non-linear responses (alternating ice thinning and recovery) as the system crossed critical thresholds. This validated the model's ability to simulate regime shifts.

5. Significance

This work represents a significant advancement in AI for Earth Science by bridging the gap between deep learning and physical domain knowledge.

Beyond Association: It moves beyond predictive correlation to identify true causal mechanisms, essential for refining climate models and predicting sea-level rise.
Data Efficiency & Robustness: By incorporating physical constraints, the model requires less data to learn robust representations and generalizes better to unseen scenarios compared to purely data-driven models.
Policy Relevance: The ability to quantify the specific causal impact of SSH on sea ice thickness provides actionable insights for understanding polar climate change and global sea-level rise mechanisms.

The authors conclude that KGCM-VAE offers a reliable, physics-informed tool for causal inference in complex, dynamic Earth systems, with future work planned to extend the framework to high-dimensional spatial fields.