Abstract
Predicting dissolved oxygen (DO) concentrations in north temperate lakes requires a comprehensive study of phenological patterns across various ecosystems, which highlights the significance of selecting phenological features and feature interactions. Process-based models are limited by partial process knowledge or oversimplified feature representations, while machine learning models face challenges in efficiently selecting relevant feature interactions for different lake types and tasks, especially under the infrequent nature of DO data collection. In this paper, we propose a Nature-Guided Cognitive Evolution (NGCE) strategy, which represents a multi-level fusion of adaptive learning with natural processes. Specifically, we utilize metabolic process-based models to generate simulated DO labels. Using these simulated labels, we implement a multi-population cognitive evolutionary search, where models, mirroring natural organisms, adaptively evolve to select relevant feature interactions within populations for different lake types and tasks. These models are not only capable of undergoing crossover and mutation mechanisms within intra-populations but also, albeit infrequently, engage in inter-population crossover. The second stage involves refining these models by retraining them with real observed labels. We have tested the performance of our NGCE strategy in predicting daily DO concentrations across a wide range of lakes in the Midwest, USA. These lakes, varying in size, depth, and trophic status, represent a broad spectrum of north temperate lakes. Our findings demonstrate that NGCE not only produces accurate predictions with few observed labels but also, through gene maps of models, reveals sophisticated phenological patterns of different lakes.
The Challenge of Predicting Dissolved Oxygen
Dissolved oxygen (DO) is often referred to as the "pulse" of a lake, serving as a critical indicator of ecosystem health and a necessity for aquatic life. However, accurately predicting DO concentrations is notoriously difficult due to the complex web of interactions involving air-water exchange, photosynthesis, respiration, and nutrient runoff.
Current modeling approaches face three primary hurdles:
- Complexity of Interactions: The sheer number of potential interactions between variables (e.g., how wind speed affects DO differently based on lake depth) makes manual selection biased and automated selection prone to noise.
- Lack of Flexibility: Many models are "global," attempting to fit a single set of rules to diverse lake types, thereby ignoring unique seasonal patterns (phenology).
- Data Scarcity: Real-world DO data is often sparse due to the manual labor required for collection, making it difficult to train data-hungry deep learning models.
The NGCE Framework: Evolution-Inspired Modeling
To address these challenges, the researchers introduced the Nature-Guided Cognitive Evolution (NGCE) strategy. At its core is the Multi-population Cognitive Evolutionary Search (MCES), an algorithm that treats feature interactions as "genomes" and predictive models as "organisms."
Multi-population Structure
Instead of a single pool of models, NGCE maintains eight distinct populations categorized by lake size (Small, Medium, Large, Extra-large) and depth layer (Surface/Epilimnion vs. Bottom/Hypolimnion). This allows models to specialize in their specific "habitat."
Evolutionary Mechanics
- Mutation: If a feature interaction is underperforming, the model replaces the mathematical operation (e.g., changing addition to multiplication) to find a better fit.
- Crossover: Models within a population share successful traits. Occasionally, models from different populations swap genetic information to promote diversity and prevent the search from getting stuck in local optima.
Two-Stage Learning Process
- Simulation Stage: To overcome data scarcity, process-based (physics-driven) models generate millions of simulated data points. The MCES evolves its feature selection within this simulated environment.
- Refinement Stage: The best-evolved models are then fine-tuned using the limited set of real-world observations, bridging the gap between simulation and reality.
Data and Experimental Setup
The study utilized a massive dataset covering 375 lakes in the Midwestern USA over a 41-year period (1979–2020). The dataset included approximately 5.58 million daily records and 39 phenological fields, encompassing lake morphometry, weather, trophic states, and land use. Data was aggregated from sources including the Water Quality Portal (WQP) and HydroLAKES.
Results and Ecological Discovery
NGCE consistently outperformed standard baselines, including LSTMs and state-of-the-art physics-informed models like EA-LSTM. In terms of Root Mean Square Error (RMSE), NGCE achieved significantly lower error rates across all lake types. For example, in "Extra-large lakes (Epi)," NGCE achieved an RMSE of 0.918, compared to 1.479 for a standard LSTM.
Beyond performance metrics, the model's "gene maps" offered significant ecological insights. By visualizing selected interactions, the authors could interpret the underlying physics. They discovered that in large lakes, DO is dominated by sediment oxygen demand, whereas small lakes are much more sensitive to local land use and weather patterns. This demonstrates that the evolutionary search successfully captured the distinct phenological patterns of different lake ecosystems.
Conclusion
The NGCE strategy represents a significant step forward in environmental modeling by fusing adaptive learning with natural processes. By mimicking evolution, the framework effectively handles complex feature interactions and data scarcity. This approach not only yields highly accurate predictions but also provides an interpretable window into the ecological dynamics of north temperate lakes, with potential applications for other environmental challenges like harmful algal bloom prediction.
Summary updated