Abstract:

Surrogate models can substantially reduce the finiteelement-analysis cost in permanent magnet motor design, but conventional parameter-based models are often tied to a fixed set of input variables. As a result, even small changes in the motor schema may require redesigning the model architecture or rebuilding the data-processing pipeline. This paper presents a schema-guided continual surrogate modeling framework based on BERT and low-rank adaptation (LoRA). Motor parameters are serialized into canonical text descriptions, allowing heterogeneous schemas to be handled by a shared language-model backbone. To support incremental design workflows, a base surrogate is first trained on the original surface permanent magnet motor schema, while newly arriving schema-expanded variants are adapted using lightweight task-specific LoRA modules and regression heads. To avoid task interference caused by sequential
LoRA adaptation, each variant uses an isolated adapter selected by a schema registry during inference. Experiments on three related motor datasets show that the proposed framework enables efficient adaptation to schema-modified designs, preserves the performance of previously learned tasks, and remains robust when input parameter clauses are reordered.