M-CaStLe: Uncovering Local Causal Structures in Multivariate Space-Time Gridded Data
| Authors | J. Jake Nichol et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2605.00398 |
| Download | |
| Categories | cs.LG, stat.ML |
Abstract
Causal graph discovery for space-time systems is challenging in high-dimensional gridded data, which often has many more grid cells than temporal observations per cell. The Causal Space-Time Stencil Learning (CaStLe) meta-algorithm was developed to address that niche under space-time locality and stationarity assumptions, but it is currently limited to univariate analyses. In this work, we present M-CaStLe. M-CaStLe generalizes the local embedding and parent-identification phases of CaStLe to jointly model local within-variable and cross-variable space-time causal structures in gridded data. Like CaStLe, by constraining candidate parents to a constant-size space-time neighborhood and pooling spatial replicates, M-CaStLe increases effective sample size to make discovery tractable in high-dimensional settings. We further decompose the resulting multivariate stencil graph into reaction and spatial graphs to aid interpretation in complex settings. We study M-CaStLe in four settings: a multivariate space-time vector autoregression benchmark with known ground truth, an advective-diffusive-reaction partial differential equation verification problem with derived physical reference structure, an atmospheric chemistry case study in a low-temporal-sample regime, and an El Niño Southern Oscillation study on reanalysis data, identifying phase-dependent ocean--atmosphere coupling. Across these settings, M-CaStLe more accurately recovers multivariate causal structure in controlled settings and identifies important physical dynamics in real-world case studies. Overall, M-CaStLe advances causal discovery for multivariate space-time systems while retaining interpretability at the grid level.
Engineering Breakdown
Plain English
This paper extends CaStLe, an algorithm for discovering causal relationships in gridded space-time data (like climate or weather measurements), to handle multivariate systems where multiple variables influence each other. The original CaStLe could only analyze one variable at a time; M-CaStLe (Multivariate CaStLe) jointly models both within-variable and cross-variable causal structures simultaneously. The core problem is that gridded data has far more spatial grid cells than time steps per cell, making traditional causal discovery statistically intractable; M-CaStLe solves this by constraining parent candidates to fixed-size local space-time neighborhoods and pooling spatial replicates to increase effective sample size. This enables discovering which variables influence which other variables across space and time in high-dimensional systems.
Core Technical Contribution
The key innovation is extending causal graph discovery from univariate to multivariate space-time systems while maintaining computational tractability in high-dimensional regimes. M-CaStLe generalizes two critical phases of the original CaStLe algorithm: the local embedding construction and parent-identification procedures now simultaneously capture both within-variable temporal dynamics and cross-variable causal influences rather than treating variables independently. The authors introduce a joint modeling approach that respects space-time locality and stationarity assumptions—critical assumptions that make the problem feasible when you have thousands of grid cells but only tens or hundreds of time steps. This is genuinely novel because prior work either handled univariate systems only, or required assumptions (like full observability or small systems) that don't apply to gridded climate, oceanographic, or geophysical data.
How It Works
M-CaStLe operates in phases: first, it constructs local embeddings for each grid cell by concatenating lagged observations from neighboring cells in space and time, creating a fixed-dimensional feature space despite the high grid dimensionality. Second, it identifies candidate parent variables for each target cell by constraining the search to a constant-size space-time neighborhood (e.g., nearby cells in a stencil pattern plus recent time steps), which reduces combinatorial explosion. Third, it uses statistical tests or regression to rank parent variables and construct a directed acyclic graph (DAG) of causal relationships, with the key trick being that it pools information across all spatial replicates (many grid cells with similar local structure) to boost sample size from potentially tens to thousands. The stationarity assumption means the same local causal patterns repeat across space, so you can treat each cell's local neighborhood as an independent sample from the same distribution. Finally, the algorithm outputs a sparse graph structure showing which variables causally influence which other variables at different spatial lags and temporal delays.
Production Impact
For engineers building systems that analyze gridded data (weather forecasting, climate modeling, oceanography, air quality monitoring), M-CaStLe provides a principled way to automatically discover causal structure instead of hand-engineering domain knowledge into models. In production, this could replace or augment manual feature engineering and physics-based priors—you'd run M-CaStLe as a preprocessing step to learn which variables and neighbors matter for your target prediction task, then use that graph structure to build more interpretable and sample-efficient downstream models. The practical benefit is significant for sparse-observation regimes: if you have 100 grid cells over 50 time steps, you'd normally have only 50 samples per grid cell, but M-CaStLe's pooling strategy treats this as 5,000 effectively independent samples for learning local causal structures. Trade-offs include computational cost (the algorithm must search and test multiple candidate parent sets), assumption sensitivity (if stationarity or locality assumptions are violated—e.g., due to climate non-stationarity or anomalous events—results degrade), and the need for sufficient temporal resolution to distinguish causal lags from noise. Integration requires careful preprocessing to align gridded data, handle missing values, and validate discovered causal structures against domain knowledge.
Limitations and When Not to Use This
M-CaStLe assumes space-time locality (causal influences only from nearby neighbors) and stationarity (same causal patterns across space and time), which can break down in real systems—seasonal shifts, climate regime changes, or long-range teleconnections in the climate system violate these assumptions. The algorithm is designed for gridded data and may not apply well to irregular observation networks, sparse sensor deployments, or systems where causal structure changes regionally or temporally. The paper also appears to be focused on discovering structure rather than measuring causal effect sizes, so while you learn "variable A influences variable B," you may not robustly estimate the strength or time-delay of that influence, limiting quantitative forecasting applications. Additionally, the abstract is truncated and doesn't provide experimental results, so it's unclear how well M-CaStLe performs compared to alternatives on benchmark datasets, what computational complexity it has, or how sensitive it is to hyperparameter choices like neighborhood size and stationarity window length.
Research Context
This work builds directly on CaStLe (Nichol et al., prior work by the same group), which demonstrated that causal discovery in high-dimensional gridded data is feasible under locality and stationarity assumptions. The multivariate extension addresses a real limitation in causal inference for geophysical systems—existing causal discovery methods (PC algorithm, FCI, constraint-based methods) don't scale to thousands of variables, and existing physics-based reduced models assume you already know the coupling structure. M-CaStLe opens the door to joint structure learning in multivariate spatio-temporal systems, which could benefit fields like climate science (discovering teleconnections between distant variables), neuroscience (multi-electrode recordings on 2D/3D arrays), and materials science (property evolution across time and space). This research direction may inspire follow-up work on relaxing the stationarity assumption, extending to non-gridded data, and integrating causal discovery with downstream prediction tasks in an end-to-end learnable framework.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::