Abstract
Scientific discovery in domains such as ecology and systems biology often requires modeling processes that vary across both time and space. While inductive process modeling has automated the discovery of temporal dynamics (ODEs), extending this to spatio-temporal dynamics (PDEs) introduces significant computational challenges. This paper introduces SCISM, a system for inducing spatial process models, and SCISM', which incorporates a transfer learning module to prune the search space. By analyzing spatial "slices" of data to learn structural constraints, SCISM' achieves a 70% reduction in search time while successfully recovering complex model structures from noisy data.
Introduction
Understanding complex environmental systems requires models that account for spatial movement—such as diffusion, advection, or sinking—alongside temporal changes. Traditional inductive process modeling (IPM) has focused on ordinary differential equations (ODEs). Transitioning to partial differential equations (PDEs) expands the search space and increases the cost of model evaluation, as each candidate must be simulated using computationally intensive numerical methods.
The SCISM Framework
The authors propose two systems to address these challenges: SCISM (Structurally Constrained Induction of Spatial Models) and its optimized counterpart, SCISM'.
SCISM: The Baseline System
SCISM extends the standard process modeling language to include partial derivatives. It employs the semi-discrete method for simulation, which discretizes spatial dimensions to transform PDEs into a system of ODEs. This allows the system to utilize established ODE solvers while representing spatial dynamics.
SCISM': Optimized Induction via MISC
SCISM' introduces the MISC (Method for Inducing Structural Constraints) module to mitigate the high cost of PDE induction. MISC utilizes a "transfer learning" approach where it first learns from simplified versions of the problem to constrain the full search.
Heuristic Search Reduction through Spatial Slicing
The core innovation in SCISM' is the "slicing" heuristic. The system assumes that if a model structure cannot explain the dynamics at a single spatial point (treated as an ODE), it is unlikely to be valid for the entire spatial volume (the PDE).
- Slicing: The system extracts data from individual spatial locations (e.g., a specific depth).
- Constraint Learning: It fits ODE models to these slices and identifies structures that consistently fail.
- Inductive Logic Programming: Using the Aleph system, it generates logical rules (constraints) that prune these failing structures from the global search space.
Experimental Evaluation
The systems were evaluated using a synthetic aquatic ecosystem dataset involving four entities (phytoplankton, zooplankton, nitrate, and detritus) and ten processes. The search library contained 24 generic processes, creating a search space of 672 possible model structures.
Accuracy and Efficiency
Both systems successfully identified the correct model structure despite the presence of Gaussian noise. However, the efficiency gains were substantial:
- Search Time: SCISM required 15.35 hours on average, whereas SCISM' completed the task in 4.73 hours.
- Pruning Power: The first spatial slice alone allowed MISC to learn eight constraints, reducing the candidate pool by 78% (from 672 to 144 structures).
Conclusion
SCISM' demonstrates that "failing fast" on simplified spatial slices can dramatically accelerate the discovery of complex spatio-temporal models. By combining numerical simulation with inductive logic programming, the system bridges the gap between raw data and interpretable scientific equations. Future work will focus on scaling these methods to multi-dimensional grids and improving numerical stability in turbulent systems.
Summary updated