Multistability and intermingledness in complex high-dimensional data

This paper introduces a computational workflow that utilizes recent advances in nonlinear dynamics to algorithmically identify multistable states in high-dimensional climate and exoplanet data, optimize key observables for monitoring, and quantify the "intermingledness" of alternative steady states and their basins of attraction.

The Gist

Imagine you are a weather forecaster, but instead of predicting tomorrow's rain, you are trying to predict the ultimate fate of a complex system like the Earth's climate.

Sometimes, nature has a "choose your own adventure" book. Depending on how you start, the story can end up in very different places, even if the rules of the game (the physics) stay exactly the same. This is called Multistability.

• Example: Imagine a ball on a hilly landscape. If you roll it from the left, it might settle in a valley on the left. If you roll it from the right, it might settle in a valley on the right. Both valleys are stable, but they are very different places. In climate science, one valley might be a "warm Earth" and the other a "Snowball Earth."

The problem is that modern climate models are so huge and complicated (like a library with millions of books) that scientists struggle to figure out:

1. How many different "valleys" (final states) actually exist?
2. Which specific measurements (like temperature or wind) tell us which valley we are in?
3. How "messy" is the boundary between these valleys? (Is it a clean line, or a tangled mess where a tiny nudge could send you to the wrong valley?)

This paper introduces a new digital detective toolkit to solve these mysteries without needing to run millions of expensive simulations manually.

The Detective's Toolkit: A Step-by-Step Analogy

Think of the researchers as detectives trying to sort a massive pile of mixed-up photos into different albums.

1. The "Feature" Filter (Simplifying the Chaos)

The raw data from these climate models is overwhelming—millions of data points per second. You can't look at all of them.

• The Analogy: Imagine you have a thousand photos of a party. Instead of looking at every single person's face, you decide to only look at three things: How loud the music is, how many people are dancing, and the average temperature of the room.
• The Paper's Method: The authors take the massive data and boil it down to a few key "features" (like averages or variations of temperature). They call this Feature Extraction.

2. The "Smart Sort" (Finding the Groups)

Now that they have simplified the data, they need to group the simulations.

• The Analogy: You throw all the photos onto a table. A smart robot (an algorithm called IA-DBSCAN) looks at the photos and says, "Hey, these 50 photos all look like 'Dancing Parties,' and these 30 look like 'Quiet Reading.'"
• The Twist: The robot doesn't guess how many groups there are. It figures it out on its own. If there are 3 distinct types of parties, it finds 3 groups. If there are 5, it finds 5.

3. The "Best Detective" Hunt (Optimization)

Sometimes, the robot picks the wrong features. Maybe "music volume" doesn't actually tell you if it's a dancing party or a reading party.

• The Analogy: The researchers run a trial-and-error loop. They ask the robot: "What if we only look at temperature? What if we only look at dancing? What if we look at both?"
• The Goal: They find the specific combination of features that makes the groups look as distinct as possible. This tells scientists: "Hey, if you want to know if the climate is collapsing, stop looking at global wind speed and start looking at North Atlantic ocean temperature."

4. The "Tangled Mess" Meter (Intermingledness)

This is the paper's most creative invention. They call it Intermingledness.

• The Analogy: Imagine two groups of people at a party: the "Dancers" and the "Readers."
  • Low Intermingledness: The dancers are all in one corner, and the readers are in another. The line between them is clear. If you nudge a dancer slightly, they stay a dancer.
  • High Intermingledness: The dancers and readers are mixed up in a giant, chaotic swirl. A dancer is standing right next to a reader. If you nudge that dancer just a tiny bit, they might accidentally bump into the reader group and become a reader!
• Why it matters: In climate science, high intermingledness means the system is unpredictable. A tiny change in the starting conditions could lead to a completely different future (a "tipping point"). This metric tells us where the system is most fragile.

Real-World Cases They Tested

The team tested their toolkit on three different "mysteries":

1. The Atlantic Ocean (AMOC): They looked at the ocean currents. They found that the ocean can exist in a "strong flow" state or a "collapsed" state. Their tool showed that temperature is the best way to tell them apart, but the boundary between these states is quite "tangled" (intermingled), meaning it's hard to predict exactly when a collapse might happen.
2. Mid-Latitude Weather: They studied wind patterns. They found that the "tangled" nature of the weather systems means that small changes can flip the climate into a different long-term pattern.
3. Alien Planets (Exoplanets): They looked at planets orbiting other stars. Here, the "groups" weren't different starting points, but different planet types (Icy, Warm, Hot). Even though this wasn't a traditional "multistability" problem, their tool helped identify which measurements best distinguish a habitable planet from a frozen one.

Why This Matters to You

• Better Early Warnings: Instead of watching every single variable in a climate model, scientists now know which specific "dials" to watch. If those dials start acting up, we know we are approaching a dangerous tipping point.
• Saving Money: Running climate simulations is expensive (like running a supercomputer for weeks). This method helps scientists stop wasting time on useless data and focus only on the numbers that actually matter.
• Understanding Risk: The "Intermingledness" meter helps us understand how fragile our climate is. If the "messiness" is high, we know that small human errors or natural fluctuations could have huge, irreversible consequences.

In a nutshell: This paper gives scientists a new pair of glasses. Instead of seeing a blurry, chaotic mess of climate data, they can now clearly see the different possible futures, know exactly which clues to look for, and understand how close we are to the edge of a cliff.

Technical Summary

1. Problem Statement

Multistability—the existence of multiple distinct steady states (attractors) for a system under the same parameters—is a critical phenomenon in complex systems like climate, power grids, and ecosystems. In climate science, this relates to "tipping points" where a system can irreversibly shift to a new state.

However, analyzing multistability in realistic, high-dimensional simulations (e.g., Earth System Models) remains poorly understood due to several challenges:

• Dimensionality: Traditional bifurcation analysis and low-dimensional techniques fail in high-dimensional state spaces.
• Subjectivity: Identifying alternative steady states currently relies on subjective visual inspection of a few representative simulations.
• Lack of Rigor: Existing methods are often not reproducible, not transferable across parameters, and cannot objectively determine which observables best distinguish between states.
• Data Constraints: Researchers often cannot run infinite initial conditions or perturbations due to computational costs, meaning they must analyze "final" datasets where re-running simulations is impossible.

2. Methodology

The authors propose a computational workflow that automates the identification, grouping, and analysis of multistable states in high-dimensional datasets. The workflow consists of the following steps:

A. Data Generation & Feature Extraction

• Input: A dataset of simulations generated from a complex model with varying initial conditions but fixed parameters.
• Dimensionality Reduction: Instead of analyzing the full spatiotemporal state, the workflow extracts diagnostic variables (e.g., global mean temperature, variance) and transforms the time series of each simulation into a feature vector.
• Feature Types: Vectors can include statistical descriptors (mean, std dev, kurtosis), temporal metrics (spectral entropy, autocorrelation), and nonlinear measures (fractal dimension, sample entropy).

B. Feature Grouping (Clustering)

• Algorithm: The workflow uses IA-DBSCAN (Iterative Advanced Density-Based Spatial Clustering of Applications with Noise).
  • Unlike $k$-means, DBSCAN does not require a pre-defined number of clusters ($k$).
  • IA-DBSCAN recursively applies clustering to subgroups to refine the separation of attractors.
  • It optimizes the clustering radius ($\epsilon$) to maximize cluster quality.

C. Optimal Feature Selection

• Optimization Loop: Since high-dimensional feature spaces are often sparse, the workflow searches for the optimal subset of features that best separate the attractors.
• Quality Metric ($q$): The authors define a grouping quality metric:
  $$q = s \cdot [w_A A - w_F F - w_L L]$$
  Where:
  • $s$: Silhouette mean (clustering quality).
  • $A$: Number of attractors found (encouraged).
  • $F$: Number of features used (penalized to encourage parsimony).
  • $L$: Number of outliers (penalized to ensure robustness against slow-converging initial conditions).
• The algorithm iterates through combinations of features to maximize $q$.

D. Intermingledness

• Definition: A new metric quantifying the degree of mixing between attractors and their basins of attraction.
• Calculation: For a given feature dimension, it calculates the ratio of the average intra-group distance (distance between points in the same attractor) to the inter-group distance (distance to the nearest other attractor).
  • $I_a \approx 1$: High intermingling (basins are mixed; small perturbations in initial conditions lead to different outcomes).
  • $I_a \approx 0$: Low intermingling (basins are well-separated; outcomes are predictable).
• Application: Calculated for both the final attractor features and the initial conditions (basins).

E. Continuation

• The workflow can track how attractors and their intermingledness change as a model parameter varies (e.g., increasing greenhouse gas forcing), using a centroid-matching algorithm to link attractors across parameter values.

3. Key Contributions

1. Objective Workflow: A fully automated, reproducible pipeline to detect multistability in high-dimensional data without prior knowledge of the number of states.
2. Intermingledness Metric: A novel diagnostic tool to quantify the "predictability" of a system's final state based on initial conditions. It identifies which observables are most sensitive to basin boundaries.
3. Feature Optimization: An algorithmic approach to determine which physical observables are most effective at distinguishing between alternative climate states, moving beyond standard metrics like global mean temperature.
4. Open Source Implementation: The workflow is implemented in Julia (using DynamicalSystems.jl), accepting NetCDF files, making it accessible for the broader scientific community.

4. Results & Case Studies

The authors validated the workflow on three diverse datasets:

Case 1: Atlantic Meridional Overturning Circulation (AMOC)

• Model: Veros (ocean model).
• Finding: Identified 5 distinct attractors (including vigorous and collapsed states) that were previously difficult to distinguish.
• Key Insight: While AMOC strength is a valid discriminator, North Atlantic surface and sub-surface temperatures were the optimal features for separation.
• Intermingledness: Basins for "collapsed" AMOC states showed higher intermingledness (lower predictability) than "vigorous" states.

Case 2: Mid-latitude Coupled Ocean-Atmosphere Flow

• Model: MAOOAM.
• Finding: Confirmed 3 distinct attractors (Low Frequency Variability modes).
• Key Insight: The optimal features for separation were different from those used in previous manual studies (specific ocean/atmosphere streamfunctions).
• Continuation: As atmospheric emissivity ($\epsilon$) increased, one attractor became increasingly intermingled with others, suggesting an approaching bifurcation or collapse.

Case 3: Exoplanet Habitability

• Model: ExoPlaSim (SAMOSA project).
• Context: This dataset represents different parameter configurations (not initial conditions), so traditional multistability does not apply.
• Application: The workflow was used to analyze the separation of "habitability regimes" (icy, warm, hot).
• Finding: Intermingledness successfully identified which diagnostics best separated the habitability classes, demonstrating the metric's utility even outside strict multistability contexts (e.g., for model intercomparison).

5. Significance and Implications

• Early Warning Systems: By identifying observables with low intermingledness (good separators) vs. high intermingledness (sensitive to noise), the method guides the placement of monitoring sensors and the development of early-warning signals for tipping points.
• Model Intercomparison: The workflow provides an objective way to compare different climate models. If models show different intermingledness patterns for the same physical variable, it may indicate model biases or structural deficiencies.
• Predictability: The metric quantifies "final state sensitivity." High intermingledness implies that minute uncertainties in initial conditions lead to drastically different climate outcomes, limiting predictability.
• Scalability: The method is computationally efficient (running in ~5 minutes for the case studies) and scales to large ensembles, making it suitable for modern Earth System Models.

6. Limitations

• Initial Condition Density: The method requires a sufficient number of initial conditions (suggested $\ge 10$ per expected attractor) to accurately map basins.
• Transient Handling: The workflow may misclassify long transients as steady states if the feature extraction window is not carefully chosen.
• Arbitrary Weights: The optimization metric ($q$) relies on empirically chosen weights ($w_A, w_F, w_L$), though the authors note the workflow remains robust to variations in these.
• Distance Metric: The intermingledness calculation depends on the choice of distance function, which can be tricky for heterotypic variables (e.g., mixing temperature and humidity).

In conclusion, this paper bridges the gap between theoretical nonlinear dynamics and practical climate modeling by providing a rigorous, automated toolkit to uncover hidden multistability and quantify the predictability of complex systems.