Machine Learning-Enhanced Feature Engineering for High-Fidelity Quasi-Isentropic Waveform Prediction under Data...

Area of Science:

• Materials Science
• Computational Physics
• Machine Learning

Background:

• Quasi-isentropic loading is crucial for determining material dynamic physical parameters and ensuring material safety.
• Acquiring these parameters often requires extensive experimental data or complex simulations, especially under data-scarce conditions.

Purpose of the Study:

• To develop a physics-guided machine learning (PGML) framework for high-fidelity prediction of quasi-isentropic loading waveforms.
• To address challenges of data scarcity in dynamic material characterization.

Main Methods:

• Integrating physical principles (shock propagation) and an attention mechanism into a machine learning model.
• Utilizing mathematically regulated inductive biases and deep feature engineering.
• Employing a 4x4 augmentation strategy for enhanced small-data prediction.

Main Results:

• Achieved R-squared > 0.96 and Mean Absolute Error (MAE) of 18.5 m/s with only 528 samples.
• Reduced shape alignment error by over 35% compared to baseline methods.
• Demonstrated breakthrough performance in accuracy and training efficiency across various impact velocities.

Conclusions:

• The PGML framework offers a data-efficient paradigm for graded structure material design.
• The approach significantly reduces reliance on resource-intensive simulations and experiments.
• SHAP analysis identified key parameters (Hill coefficient and curvature modulation parameter) influencing waveform modulation, providing interpretable insights.