Ground Vibration Propagation Model Improves Accuracy with Machine Learning Integration

Research at the Beijing High Energy Source site characterised ground vibration propagation between 1 and 100 Hz. A hybrid iterative fitting method, combining machine learning with the Bornitz formula, accurately models these vibrations with improved accuracy and interpretability compared to existing empirical and ‘black box’ approaches.

Understanding and mitigating ground vibrations is critical for facilities housing sensitive equipment, particularly large-scale scientific instruments. External disturbances, even seemingly minor ones, can compromise data quality and operational precision. Researchers at Tsinghua University and the Chinese Academy of Sciences have investigated these effects at the Beijing High Energy Photon Source (HEPS) site, a major accelerator facility. In a study detailed in their paper, ‘Propagation and Attenuation Characteristics of Ground Vibrations: Beijing High Energy Photon Source (HEPS) Experiments and Intelligent Law Discovery’, Pei-Yao Chen, Chen Wang, Jian-Sheng Fan (all from Tsinghua University’s Department of Civil Engineering) and Fang Yan, Chao-Yang Zhang, Xiang-Yu Tan, and Guo-Ping Lin (from the Institute of High Energy Physics, Chinese Academy of Sciences) present a novel hybrid approach combining experimental data, machine learning, and established geophysical modelling to characterise and predict ground vibration behaviour.

Hybrid Modelling Accurately Predicts Ground Vibration Propagation

Accurate characterisation of ground vibration propagation and attenuation is vital for protecting sensitive instrumentation, particularly at large-scale facilities such as the Beijing High Energy Source (HEPS). Researchers have demonstrated a novel hybrid methodology combining machine learning with established physical modelling to achieve accurate predictions and interpretable results. This approach offers a robust framework for characterising ground vibration behaviour, with potential applications extending beyond the HEPS site.

The team conducted swept-frequency excitation tests at the HEPS facility, employing an electrodynamic vibrator to generate controlled vibrations across a 1-100 Hz frequency range. The resulting empirical data served to refine their models. Crucially, the research does not rely on purely ‘black-box’ machine learning; instead, it leverages machine learning techniques to augment the established Bornitz formula – an empirical equation used to estimate the transmission of ground-borne vibration.

Probabilistic analysis was employed to quantify the uncertainty associated with the derived formula, measuring its reliability and enabling more confident predictions under varying conditions. The researchers also elucidated the formula’s physical rationale, enhancing its interpretability and demonstrating a clear link between the model and the underlying physics. This emphasis on interpretability distinguishes the approach from purely data-driven methods, which often lack explanatory power.

The core innovation lies in integrating data-driven insights to refine and improve existing physical models, rather than replacing them. By building upon the Bornitz formula, the team balanced predictive accuracy and physical plausibility, resulting in a more robust and reliable model.

Comparative analysis revealed this hybrid method’s superiority over empirical formulas from previous studies and conventional black-box machine learning models, demonstrating improvements in accuracy and interpretability. This enhancement facilitates more effective mitigation strategies and improved protection of sensitive instruments.

The methodology’s strength resides in its ability to balance predictive power with physical rationality, ensuring that model outputs reflect underlying physical processes, rather than merely statistical correlations. Probabilistic analysis quantifies the uncertainty within the derived formula, while the foundation in established physics ensures the model’s outputs are meaningful and reliable.

This successful integration of machine learning and established physics represents an advancement in the field, offering a pathway for discovering propagation and attenuation laws applicable beyond the HEPS facility. The research provides a valuable template for future investigations seeking to combine data-driven techniques with established scientific principles to address complex engineering challenges.