Physics and artificial intelligence collaborative evaluation-prediction framework for rapid underwater acoustic transmission loss modeling

Artificial intelligence (AI) has recently attracted considerable attention for low-cost acoustic field prediction. However, existing AI methods struggle with uncontrollable prediction reliability across diverse marine environments and lack confidence evaluation mechanisms, hindering their direct application in practical engineering. To address these issues, this paper proposes a physics-AI collaborative evaluation-prediction framework. Inspired by uncertainty assessment, we construct an AI prediction quality evaluation module to conduct environmental confidence assessments and automate task routing: acceptable samples are assigned to AI for rapid prediction, while unacceptable ones revert to physical models to ensure accuracy. Additionally, principal component analysis is utilized to linearly compress the high-dimensional acoustic field, further alleviating the model's computational and storage burdens. Validated on Bellhop-generated data using fixed bottom parameters, experiments demonstrate that the evaluation model identifies acceptable samples with 94.38% precision. By effectively isolating acceptable predictions [average root mean square error (RMSE) 2.614 dB] from unacceptable ones (average RMSE 4.508 dB), overall computation time is reduced by over 50%. Notably, this flexible framework requires only lightweight fine-tuning to adapt to existing neural networks, providing a viable paradigm for applying intelligent underwater acoustic models in practical engineering.