GeoEvolve: Guided Evolution in Geospatial Modeling

• GeoEvolve is a multi-agent, knowledge-guided evolutionary framework that integrates geospatial theory with code evolution to automate algorithm discovery.
• It employs a dual-loop architecture where an inner loop evolves candidate algorithms and an outer loop refines them using explicit theoretical priors and performance metrics.
• Empirical evaluations show significant improvements in spatial interpolation and uncertainty quantification, validating its practical impact on geospatial modeling.

GeoEvolve is a multi-agent, knowledge-guided evolutionary framework for automated discovery and refinement of geospatial algorithms. Designed to address the unique challenges of geospatial modeling—such as spatial autocorrelation, heterogeneity, and the necessity of theoretical coherence—GeoEvolve combines evolutionary code search with retrieval-augmented geospatial knowledge integration. The system has demonstrated empirical improvements over classical and baseline LLM-driven approaches, particularly in spatial interpolation and geospatial conformal prediction, and incorporates explicit mechanisms to ensure the injection of structured domain knowledge at all stages of the algorithmic search process (Luo et al., 25 Sep 2025).

1. Motivation and Conceptual Framework

Geospatial analysis poses algorithm discovery challenges not seen in generic computational domains. Key sources of complexity include multiscale spatial autocorrelation, nonstationarity, multimodal data (maps, networks, and imagery), and frequent reliance on domain-specific operators and priors (e.g., variogram modeling in geostatistics). While LLM-based frameworks such as AlphaEvolve or OpenEvolve can explore a wide class of code mutations, they tend to produce solutions that lack geospatial theoretical validity when guided only by generic loss metrics.

GeoEvolve directly addresses this by embedding geospatial knowledge into the evolutionary loop. A retrieval-augmented generation (GeoKnowRAG) module injects explicit theoretical priors, algorithmic templates, and critical implementation heuristics into every stage of code search, systematically steering the system away from statistically or geographically misinformed algorithmic drift (Luo et al., 25 Sep 2025).

2. System Architecture and Workflow

GeoEvolve is structured as a dual-loop, multi-agent system:

• Inner Loop (Code Evolver Agent):
  • Operates on the OpenEvolve code base.
  • Generates, mutates, and tests candidate algorithms, evaluating them according to task-specific loss functions, e.g., root mean squared error (RMSE) for kriging, and interval score for spatial uncertainty quantification.
  • Propagates high-performing ("elite") candidates across iterations:

  $$A^* = \arg\min_{A\in\mathcal{A}} \mathcal{L}(A; \mathcal{D})$$

  where $\mathcal{A}$ denotes code search space and $\mathcal{L}$ is the loss over data $\mathcal{D}$.

• Outer Loop (Agentic Controller and GeoKnowRAG):

  • Aggregates local elites from inner loop.
  • Maintains a persistent set of global elites to ensure monotonic improvement.
  • Performs semantic analysis via the Code Analyzer agent, which inspects best-performing code to diagnose missing geospatial operators or theoretical gaps.
  • Queries the GeoKnowRAG module to retrieve relevant geospatial knowledge, including formulas, best practices, and evaluation diagnostics, and encodes this into prompts via the Prompt Generator.
  • The enriched prompts, containing explicit theoretical and implementation constraints, are injected into subsequent inner loop iterations.

Communication and information flow are represented as: Code Evolver → (code+metrics) → Code Analyzer → (queries) → GeoKnowRAG → (texts & formulas) → Prompt Generator → (prompt) → Code Evolver.

The top-level pseudocode is: $\mathcal{L}$5 (Luo et al., 25 Sep 2025)

3. Structured Domain Knowledge: GeoKnowRAG

GeoKnowRAG is the knowledge retrieval module underpinning GeoEvolve’s theoretical grounding. The knowledge base comprises five topical categories—geostatistics, spatial theory, GIScience, spatial statistics, spatial modeling—derived from ~141 curated documents (Wikipedia, arXiv, GitHub). Documents are chunked and embedded using OpenAI’s text-embedding models, then queried using multi-angle expansion and Reciprocal Rank Fusion (RRF) for robust retrieval across theory, implementation, and evaluation sub-domains.

The system can inject domain heuristics (e.g., Silverman’s rule for bin widths), explicit mathematical operators (e.g., Matérn, exponential, Gaussian variogram families), and diagnostic criteria (e.g., model selection via AIC/BIC). Domain-guided mutations are explicitly informed by retrieved formulas and patterns, with prompt generation encoding these insights for inner-loop code evolution.

A typical retrieval and injection sequence for variogram model selection:

• Analyzer requests: "variogram model selection via AIC/BIC"
• GeoKnowRAG returns AIC/BIC formulas and definitions for variogram families (Exponential, Gaussian, Linear, Matérn)
• Prompt generator instructs evolver: "implement four variogram families...select by min AIC/BIC" (Luo et al., 25 Sep 2025)

4. Core Algorithms and Extensions

GeoEvolve’s evolution framework is applied to canonical spatial tasks:

4.1 Spatial Interpolation (Kriging)

• Model: $Z(s) = \mu + \epsilon(s)$ with second-order stationarity $\mathrm{Cov}[\epsilon(s), \epsilon(s+h)] = C(h)$
• Kriging estimator: $\hat Z(s_0) = \sum_{i=1}^n \lambda_i Z(s_i)$, $\sum \lambda_i = 1$
• Matrix form: $K\lambda = k$ with $K_{ij} = C(\|s_i - s_j\|)$, $\mathcal{A}$0
• Evolution over variogram family:

$\mathcal{A}$1

with $\mathcal{A}$2 (exponential), $\mathcal{A}$3 (Gaussian), $\mathcal{A}$4 (Matérn).

• Model selection and fitting leverages AIC/BIC and robust (L1/weighted least squares) metrics, adaptive binning heuristics, and regularization schemes (e.g., K-NN localization) (Luo et al., 25 Sep 2025).

4.2 Spatial Uncertainty Quantification (GeoCP)

• Weighted conformal prediction intervals leveraging geodesic kernel weights:

$\mathcal{A}$5

e.g., $\mathcal{A}$6

• Weighted empirical quantile for confidence region:

$\mathcal{A}$7

yielding interval $\mathcal{A}$8.

• Interval score for evaluation:

$\mathcal{A}$9

GeoEvolve also introduces novel algorithmic components, including automatic variogram-family selection, multi-start fitting strategies, robust binning procedures, condition-based localization, and adaptive data transformations (Luo et al., 25 Sep 2025).

5. Benchmarks and Empirical Evaluation

GeoEvolve has been empirically validated on two canonical geospatial tasks:

• Spatial interpolation (Ordinary Kriging on Australian geochemistry):
  • GeoEvolve-Kriging reduces RMSE by $\mathcal{L}$0–$\mathcal{L}$1 relative to OpenEvolve, and $\mathcal{L}$2–$\mathcal{L}$3 relative to original kriging.
  • Error-map visualizations indicate improved surface accuracy and reduced overfitting.
• Spatial uncertainty quantification (GeoCP on real estate):
  • Interval Score reduced by $\mathcal{L}$4 compared to classical conformal prediction, with uncertainty maps capturing genuine spatial heterogeneity.

Ablation studies show that removing the GeoKnowRAG guidance degrades performance (positioning GeoEvolve between unguided OpenEvolve and the fully guided variant), and that naive prompting with generic geospatial cues yields minor or inconsistent gains (Luo et al., 25 Sep 2025).

6. Significance and Limitations

Theory-driven code evolution, as instantiated in GeoEvolve, reduces the effective algorithm search space and prioritizes physically and statistically meaningful mutations, resulting in gains in accuracy and interpretability. The explicit decomposition of search, critique, retrieval, and prompt construction across agents enables iterative, multi-faceted refinement. The nested dual-loop structure supports both local exploitation and global, theory-guided exploration (Luo et al., 25 Sep 2025).

Limitations include the current size of the knowledge base (141 sources), the computational expense of nested LLM-in-the-loop iterations, and a need for further extensions to multimodal geospatial inputs (e.g., raster imagery, complex networks). Planned directions involve expanding retrieval corpora, supporting a broader class of geospatial tasks (classification, clustering, network analysis), and deeper integration with open-source LLMs.

GeoEvolve offers a scalable reference implementation for automated, theory-guided geospatial model discovery and is positioned as a foundation for future AI-for-Science systems in spatial analysis (Luo et al., 25 Sep 2025).