AI LAB / RESEARCH WORKSPACE
From observation to insight.
Explore real series, test your assumptions and build a clearer basis for decisions.
Discover patterns. Test predictions.
Build a multivariable table, explore relationships and evaluate a selected target.
Start with matching data
Each click adds up to three unseen examples from the database pool. All 2–6-variable candidate spaces are indexed; shared observations are checked progressively and verified combinations are stored. New data creates a fresh index. Automatic examples require at least 30 complete observations and 10 per variable as a screening heuristic, not a guarantee of reliability. Findings remain exploratory.
Select 2–6 variables. Rows are matched by exact location code and period; values are not automatically aggregated or filled.
CSV includes all periods in the joined table; missing values stay blank and complete_row marks complete rows. Download JSON for detailed source and method information.
Select variables to build the analysis table.
Methods and interpretation
Pearson correlations use pairwise complete rows and display each pair’s sample size. Shared trends can inflate correlation; no causal claim or p-value is made. Clustering and PCA use complete rows, exclude constant columns and standardize each variable separately. Cluster counts are user-selected.
Ridge predicts the selected target one observation ahead, using previous values of all selected variables, including the target. Penalties 0.1, 1, 10 and 100 are compared in three expanding training windows; a final window is excluded from selection. Scaling and coefficients use training only. Within each window, weights remain fixed; previous observations become available for the next one-step prediction. Publication delays are not modelled, so this is retrospective evaluation, not a live forecast.
At least 20 complete rows for multivariable exploration, 48 consecutive complete rows for ML; maximum 1,200 complete rows for these methods. No imputation. Duplicate period cells are excluded. Inspect location names and period definitions before interpretation. Changing settings after inspecting the test is exploratory. Random Forest is not part of this version.
Data analysis and visualization catalogue
Explore imported data, see applicable methods and inspect their limits.
Applicable methods and limitations
Proposed scientific uses — not implemented
Click a sea cell for its available time series. Smoothing is visual only.
Values and quality
- 01Choose your evidenceInstitution, series and location
- 02Shape the analysisMethod and stated assumptions
- 03Read & share findingsEvidence, limits and next steps
Daily Sea Surface Temperature (SST) & Marine Heatwave Pulse
Interactive 45-year climate spaghetti trajectories, 1991–2020 climatology baseline, thermal anomalies, and verified 4D spatial raster layers.
Loading verified data…
Method, scope and sources
These are GOMD calculations from NOAA OISST. They are not reproductions of Mercator’s published numerical results. Regional boxes are not official basin or maritime boundaries. Surface only: depth is not animated.
Reference chart types · Mercator Ocean · Hobday MHWAnalysis canvas
Your observations, scenarios and decision context.
Let the evidence take shape.
Choose an institution, a series and a location. Run the model to explore the trend and see what your assumptions change.
A scenario describes what follows from an assumption. Review the uncertainty and limits before using a result in a decision.