Uncertainty aware machine learning for bridging simulation and experiment in high throughput materials...

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning Applications

Background:

• High-throughput materials characterization accelerates materials discovery.
• Machine learning (ML) is crucial for high-throughput characterization but faces challenges with limited labeled experimental data and prediction uncertainty.
• Adapting ML models from simulation to real-world experiments is difficult due to data discrepancies.

Purpose of the Study:

• To develop a novel approach for adapting ML models trained on simulation data to analyze experimental data, even with limited labels.
• To incorporate uncertainty estimation into ML models for more reliable predictions in experimental settings.
• To address the scarcity of labeled experimental data in ML-driven materials discovery.

Main Methods:

• Developed Uncertainty-aware Simulation-to-Experiment Modeling (USEM).
• Employed adversarial domain adaptation in the latent space to bridge simulation-experiment data gaps.
• Integrated spectral-normalized neural Gaussian processes (SNGP) for predictive uncertainty quantification.

Main Results:

• Demonstrated effective adaptation of ML models from simulation to experimental X-ray diffraction (XRD) data.
• Achieved improved predictive accuracy on experimental data.
• Successfully identified out-of-distribution samples, enhancing model trustworthiness.
• Showcased the ability to analyze unlabeled or sparsely labeled experimental data.

Conclusions:

• USEM provides a scalable and trustworthy solution for high-throughput characterization in ML-driven materials discovery.
• The approach enhances the practical application of ML in experimental materials science by overcoming data limitations and uncertainty issues.
• USEM facilitates more reliable and efficient materials discovery through improved ML model performance on experimental data.