Answers are grounded in this paper and cite their sources.
Wave-based inverse problems, which determine material or structural parameters from measured wavefields, are fundamental to non-destructive evaluation (NDE), biomedical diagnostics, and geophysics. For repeated-query applications, deep learning shifts the computational effort from iterative inversion to offline training, enabling low-latency prediction, yet large-scale adoption remains limited by challenges in reproducibility, fair architecture comparison, large-dataset handling, and stable distributed training. This paper introduces WaveDL, an open-source Python framework that addresses these challenges through an integrated pipeline covering memory-efficient data streaming, a model zoo comprising 71 network variants adapted for multi-target regression, synchronized multi-GPU training, soft physics-constraint penalties, and automated deployment export via Open Neural Network Exchange (ONNX). Built on PyTorch and Hugging Face Accelerate, WaveDL enables a complete workflow from raw data to deployed model with a single command, lowering the barrier for domain scientists to apply deep learning without specialized machine learning expertise. The framework is demonstrated using a 120 GB Lamb wave dispersion curve dataset to evaluate out-of-core training and systematic cross-architecture benchmarking for finite-dimensional parameter inversion. WaveDL is freely installable via pip install wavedl . PROGRAM SUMMARY Program Title: WaveDL CPC Library link to program files: (to be assigned) Developer’s repository link: https://github.com/ductho-le/WaveDL Licensing provisions: MIT License Programming language: Python 3.11+ Nature of problem: Inferring material or structural parameters from measured wavefields poses a challenging inverse problem that can be time-consuming and computationally expensive to solve with classical iterative methods. Deep learning offers an appealing alternative, but practical adoption often requires separate tools for reproducible training, out-of-core data handling, coordinated multi-GPU execution, soft physics-constraint penalties, and export to industrial environments. Solution method: WaveDL addresses each challenge through a dedicated subsystem: large-scale datasets are handled via memory-mapped streaming with near-constant input-memory overhead independent of input dataset size (aside from compact target storage); multi-GPU training is coordinated through synchronization of early stopping and metric aggregation; physical-law violations are penalized through soft loss terms parsed from user-supplied expressions or Python files; and trained models are exported to ONNX with embedded denormalization for direct use in MATLAB, LabVIEW, and C++ toolchains. Restrictions: Training acceleration is supported on NVIDIA (CUDA) GPUs and Apple Silicon (MPS); CPU-only execution is available but substantially slower. Memory-mapped loading requires datasets in NPZ, HDF5, or MAT v7.3 format.
No takes yet. Share an insight, caveat, or question.
Le et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: