AOT-POT: Adaptive Operator Transformation for Large-Scale PDE Pre-training

This paper addresses the fundamental challenge of pre-training neural operators on diverse PDE datasets for building general-purpose surrogate models. Existing methods primarily tackle this by increasing model capacity while leaving target solution operators unchanged. The proposed Adaptive Operator Transformation (AOT) method adapts the target solution operators themselves, providing a new paradigm for multi-PDE pre-training that handles the inherent complexity and structural diversity of PDE solution operators.