Abstract Landscape evolution model (LEM) is an effective numerical simulation tool to predict topographic changes in response to tectonic uplift and surface erosion. Combining with an inverse algorithm, the LEM have been used for deducing the tectonic and climatic factors from the current landforms. However, the predictive capability of LEMs is fundamentally limited by uncertainty in boundary conditions and parameters, particularly the spatially variable rock uplift rate field, which serves as the primary driving force in landscape evolution. This study introduces a novel, data‐driven method to invert for the initial uplift field using a genetic algorithm (GA). By coupling a GA with the LEM Fastscape and employing present‐day topography as the target, our method iteratively optimizes candidate uplift rate fields. Key features of this work include a multi‐dimensional fitness function integrating traditional geomorphic metrics with perceptual similarity (LPIPS) to better capture realistic landscape features, and an effective dimensionality reduction strategy for parameterizing the uplift field, enabling efficient inversion of complex spatial patterns using a GA. The method's performance is rigorously evaluated through a series of synthetic experiments with progressively complex uplift patterns, demonstrating the robust performance for simple Unimodal and medium complexity bimodal uplift patterns and a modest degradation for Complex Sinusoidal Patterns. Applying the method to the tectonically active Taiwan Central Range, our method effectively converges to an optimal uplift rate field that aligns with independent thermochronological constraints. This GA‐based inversion provides a robust and objective approach for determining crucial LEM initial conditions, enhancing our ability to decipher the interactions between tectonic and Earth surface processes in tectonically active regions.
Zhao et al. (Sun,) studied this question.