| About this Abstract |
| Meeting |
MS&T24: Materials Science & Technology
|
| Symposium
|
Uncertainty Quantification Applications in Materials and Engineering
|
| Presentation Title |
Uncertainty Quantification of Material Properties in Data-Poor Regimes Using Transfer Learning and Gaussian Process Regression |
| Author(s) |
Sara Akhavan Abdollahian, Soumya Sridar, Wei Xiong, Hessam Babaee |
| On-Site Speaker (Planned) |
Sara Akhavan Abdollahian |
| Abstract Scope |
Quantifying the effect of uncertainties in processing parameters on the material properties is of vital importance in materials science. Performing uncertainty quantification (UQ) requires building an accurate surrogate model between the input variables and the quantities of interest (QoIs). However, building such surrogate models for material properties is challenging since the input space is high dimensional and there is a lack of sufficient high-fidelity data to train such models. This is due to the prohibitive cost of high-fidelity data collection or performing high-fidelity simulations. Hence, we present a transfer learning (TL) approach based on Gaussian process (GP) regression, referred to as TL-GP. In TL-GP, the kernel is trained using data generated using the CALPHAD (Calculation of Phase Diagrams) approach. The trained kernel is then utilized in GP predictions conditioned on a few high-fidelity experimental data. To demonstrate the efficiency of this UQ approach, ZrB2 is chosen as the candidate material. |