@Article{fdmp.2026.087218,
AUTHOR = {Zongzheng Jiao, Shuaikang Yang, Junlong Yin, Shaohui Wang, Minpo Jung},
TITLE = {Benchmarking Physical-Parameter Conditioning Strategies for Data-Driven Hydro-Mechanical Field Forecasting},
JOURNAL = {Fluid Dynamics \& Materials Processing},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/fdmp/online/detail/28094},
ISSN = {1555-2578},
ABSTRACT = {Hydro-mechanical (HM) simulations of porous-media systems—such as those used in geotechnical engineering, groundwater flow, formation consolidation, and underground-structure safety assessment—become computationally expensive when large material- and load-parameter spaces must be explored for design optimization, uncertainty quantification, or real-time decision support. Although data-driven surrogate models can accelerate such analyses, it remains unclear whether explicitly conditioning a history-based predictor on physical parameters offers a meaningful advantage over learning directly from the temporal evolution of the physical fields. This study systematically benchmarks five physical-parameter conditioning strategies—token concatenation, feature-wise linear modulation (FiLM), weak FiLM, adaptive instance normalization (AdaIN), and gated residual conditioning (GRC)—implemented within a common convolutional neural network (CNN)–Transformer backbone. The evaluation uses a 108-case OpenGeoSys consolidation dataset and a unified protocol comprising a strict case-level split, five independent training repetitions, random-condition and zero-condition controls, an α-ablation study, and equal-budget, mechanism-specific hyperparameter tuning. The results show that the temporal evolution of the field variables contains most of the information required for forecasting, while the physical parameters primarily provide an auxiliary correction signal. Among the explicitly conditioned models, GRC provides the best overall combination of predictive accuracy and robustness. Under the shared training protocol, however, the unconditional CNN-Transformer slightly outperforms all conditioned variants, while the long short-term memory (LSTM) baseline achieves the lowest overall prediction error. These findings are specific to the present fixed-node, single-geometry benchmark and should not be interpreted as a universal ranking of surrogate-model architectures.},
DOI = {10.32604/fdmp.2026.087218}
}