Introduction
The Spallation Neutron Source (SNS) at Oak Ridge National Laboratory (ORNL) serves as a premier facility for neutron scattering research. At its core, high-energy proton pulses strike a liquid mercury target to produce neutrons. However, this process subjects the target vessel to extreme conditions. Each pulse deposits significant energy in microseconds, inducing intense pressure waves and cavitation—the formation and collapse of micro-bubbles—that erode the stainless-steel containment.
Predicting the structural integrity and lifespan of these targets requires accurate constitutive models. The current standard, the Riemer model, is computationally intensive, requiring approximately 25 hours on 96 processors for a single high-fidelity simulation. This makes traditional manual calibration of physical parameters, such as mercury density and the speed of sound, impractical.
Methods
To overcome the computational bottleneck, researchers developed two surrogate-based calibration frameworks:
- Evolutionary Neural Calibration (ENC): This approach utilizes a Feedforward Neural Network (FNN) to act as a surrogate for the expensive Sierra/SolidMechanics simulation code. By training the FNN to predict strain results in milliseconds, the team could employ evolutionary algorithms—including Genetic Algorithms, Grey Wolf Optimizer, and Differential Evolution—to search for the optimal physical parameters that match experimental sensor data.
- Sparse Polynomial Expansions (SPE): This method focuses on modeling the discrepancy between simulation and experiment. Using compressed sensing, the researchers constructed a mathematically efficient polynomial to describe the error landscape. Global optimization was then performed using Directional Gaussian Smoothing (DGS).
Datasets and Experimental Setup
The calibration relied on a robust dataset comprising:
- Experimental Data: Multi-year strain measurements from 12 sensors mounted on the exterior of the mercury target vessel.
- Simulation Data: A library of 603 high-fidelity simulations generated via the Sierra/SolidMechanics code.
- Data Partitioning: The surrogate models were trained on 80% of the simulation data (482 samples) and validated on the remaining 20% (121 samples).
Key Insights and Results
The application of these machine learning frameworks yielded significant improvements in predictive accuracy and provided deep physical insights:
- Accuracy Improvements: Calibrated simulations showed a 7% average improvement in signal prediction accuracy. Mean Absolute Error (MAE) across all sensors was reduced by 8%, with some individual sensors seeing improvements of up to 30%.
- Reliability: The calibrated model achieved an 84% success rate in falling within experimental confidence bounds, up from 77% with previous best-guess parameters.
- Diagnostic Machine Learning: A critical finding was that the optimal parameters identified by the models were often "nonphysical" for pure mercury (e.g., lower-than-expected density). This suggests that the ML models are identifying deficiencies in the underlying Equation of State (EOS). Specifically, the current models do not fully account for the complex interaction between gas bubbles and liquid mercury.
- Methodological Convergence: Both the ENC and SPE approaches arrived at nearly identical parameter sets, providing strong cross-validation for the results.
Conclusion and Future Work
This study demonstrates that machine learning surrogates can effectively calibrate complex, high-fidelity physics simulations. The "nonphysical" results have served as a diagnostic tool, highlighting the need for more sophisticated physical models. Future research will focus on implementing a Bayesian framework for better uncertainty quantification and developing two-phase flow models to explicitly capture bubble dynamics within the mercury target.
Summary updated