Ozge Ozbayram
Ozge Ozbayram has a total of 9 co-authors. Here are the top co-authors by number of shared publications:
- Audrey Olivier
Univ. of Southern California
2 shared publications - Lori Graham-Brady
Department of Civil and Systems Engineering, Johns Hopkins University 1 , Baltimore, Maryland 21218
2 shared publications Aditya Venkatraman
1 shared publicationAndreas E. Robertson
1 shared publicationDaniel Olsen
1 shared publicationLori Graham-Brady
1 shared publication
Latest Publications
- Open Access
- 17 March 2026
- Published by Elsevier in SSRN Electronic Journal
Active Learning with Heteroscedastic Gaussian Process Regression Model
Active learning aims to minimize data requirements for accurate surrogate modeling; yet when noise varies across the input space, both the model and sampling strategy must account for heteroscedasticity, which existing approaches often neglect. We present an enhanced Heteroscedastic Gaussian Process Regression (HGPR) framework with active learning for input-dependent noise. First, we introduce an enhancement of an existing HGPR model, leveraging L1 regularization to learn a sparse model for the aleatoric noise. This further enables automatic selection of relevant input features and polynomial degrees within a single optimization run, eliminating computationally expensive cross-validation. Secondly, we study active learning in this heteroscedastic setting. We adapt and evaluate four acquisition functions (Expected Improvement for Global Fit (EIGF), Active Learning Cohn (ALC), Mutual Information (MI), and Fisher Information (FI)), revealing distinct sampling behaviors. EIGF oversamples high-noise regions, degrading uncertainty calibration; FI clusters samples to refine noise parameters at the cost of mean prediction accuracy; while MI and ALC naturally balance exploration with noise-aware exploitation. Leveraging these complementary behaviors, we propose a hybrid strategy that transitions from EIGF to FI, achieving calibrated uncertainty with interpretable two-phase sampling. Experiments on synthetic functions and a microstructure-property prediction problem show that MI achieves the best overall performance, while the hybrid approach offers comparable accuracy.
- 17 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessActive Learning with Heteroscedastic Gaussian Process Regression Model
Active learning aims to minimize data requirements for accurate surrogate modeling; yet when noise varies across the input space, both the model and sampling strategy must account for heteroscedasticity, which existing approaches often neglect. We present an enhanced Heteroscedastic Gaussian Process Regression (HGPR) framework with active learning for input-dependent noise. First, we introduce an enhancement of an existing HGPR model, leveraging L1 regularization to learn a sparse model for the aleatoric noise. This further enables automatic selection of relevant input features and polynomial degrees within a single optimization run, eliminating computationally expensive cross-validation. Secondly, we study active learning in this heteroscedastic setting. We adapt and evaluate four acquisition functions (Expected Improvement for Global Fit (EIGF), Active Learning Cohn (ALC), Mutual Information (MI), and Fisher Information (FI)), revealing distinct sampling behaviors. EIGF oversamples high-noise regions, degrading uncertainty calibration; FI clusters samples to refine noise parameters at the cost of mean prediction accuracy; while MI and ALC naturally balance exploration with noise-aware exploitation. Leveraging these complementary behaviors, we propose a hybrid strategy that transitions from EIGF to FI, achieving calibrated uncertainty with interpretable two-phase sampling. Experiments on synthetic functions and a microstructure-property prediction problem show that MI achieves the best overall performance, while the hybrid approach offers comparable accuracy.
- 2 March 2025
- Published by Elsevier in Mechanics of Materials
- Vol. 205
Batch active learning for microstructure–property relations in energetic materials
- 2 March 2025
- Published by Elsevier in Mechanics of Materials
- Vol. 205
Batch active learning for microstructure–property relations in energetic materials
- 26 August 2024
- Published by Elsevier in Computer Methods in Applied Mechanics and Engineering
- Vol. 431
Heteroscedastic Gaussian Process Regression for material structure–property relationship modeling
- 26 August 2024
- Published by Elsevier in Computer Methods in Applied Mechanics and Engineering
- Vol. 431
Heteroscedastic Gaussian Process Regression for material structure–property relationship modeling