Location-conditioned generation of tailored wave sequences using a transformer-based inverse model

The inverse wave generation problem of inferring wavemaker motion from wave elevation records at arbitrary locations in a wave flume is addressed. This inverse mapping is inherently ill-posed due to dispersive wave propagation, nonlinear wave–wave interactions, and the loss of spatial information from localized observations. To address these challenges, a data-driven and location-conditioned inverse modeling framework based on a Transformer encoder with a mixture-of-experts regression head is proposed. The model is trained on a large dataset generated using a high-fidelity, open-source numerical simulation tool covering a wide range of sea states and probe locations. Model performance is evaluated using the surface similarity parameter under both wave phase-level and sea state-level data splitting strategies. Performance of the model is analyzed from various perspectives such as dependence on the sea states and probe location. Spectral analysis of the predictions and ablation study provide further insights into model’s performance and architecture. The results show that the proposed approach achieves acceptable or high accuracy for the majority of sea states. The probe-wise analysis indicates a mild degradation in performance with increasing distance from the wavemaker. The ablation study confirms the importance of the mixture-of-experts structure and explicit location conditioning for robust inverse predictions.