Research
My research is rooted in the area of Computer Experiments/Uncertainty Quantification/Digital Twins, but spans a range of applications from COVID-19 analysis to deep learning.
Computer model simulations have become increasingly important for studying complex phenomena, such as climate change, response to natural crises, and the spread of pandemic diseases. Despite the popular use of computer simulations, they often encounter three challenges in practice:
- many computer simulations are prohibitively expensive, which require conducting an experiment by which a cheaper, accurate statistical emulator needs to be developed to approximate the computer simulations;
- computer models may be imperfect, so in order to make meaningful predictions, it is essential to calibrate a computer model by aligning the model’s response with physical experimental data such that it can accurately reflect the real-world system;
- optimizing computer model outputs that are expensive to simulate can be almost impossible.
My research focus is to tackle these challenges by providing modern, efficient statistical approaches.
Highlighted Research Contributions
Multi-Fidelity or Single-Fidelity?
In many scientific applications, researchers combine expensive high-accuracy simulations with cheaper low-accuracy ones. But is multi-fidelity modeling always better than using a single high-fidelity simulator?
In this work (Sung et al., 2024), we provide a theoretical understanding of when multi-fidelity emulation improves prediction accuracy — and when it does not. We show that blindly adding low-fidelity data can sometimes hurt performance, and we give guidance for designing computer experiments efficiently.
Sung, C.-L., Ji, Y., Mak, S., Wang, W., & Tang, T. (2024). SIAM/ASA J. Uncertainty Quantification, 12(1), 157–181.
Beyond Linear Relationships in Multi-Fidelity Modeling
Traditional multi-fidelity models assume a simple linear relationship between low- and high-fidelity simulations. But real-world systems are rarely that simple.
In this work (Heo and Sung, 2025), we introduce the RNA (Recursive Non-Additive) model, allowing flexible nonlinear dependence across fidelities while retaining closed-form expressions for the posterior mean and variance. This preserves the analytical tractability of Gaussian processes and improves predictive accuracy in complex applications.
Open-source implementation: RNAmf (CRAN).
Heo, J., & Sung, C.-L. (2025). Technometrics, 67(1), 58-72.
Regression with Functional Inputs: When the Input is a Curve
In many scientific problems, inputs are not just numbers — they can be entire functions. For example, instead of a scalar parameter, the input might be a curve such as $\sin(x)$, $\cos(x)$, or a spatial profile describing material properties.
In this work (Sung et al., 2024), we develop a new class of Gaussian process models that can handle functional inputs directly. We introduce kernel functions designed for inputs that are curves or functions, rather than finite-dimensional vectors.
Open-source implementation: Reproducibility (GitHub).
Sung, C.-L., Wang, W., Cakoni, F., Harris, I., & Hung, Y. (2024). Statistica Sinica, 34(4), 1883-1902.
Calibration with Heteroscedastic Measurement Errors
Computer models are widely used to represent real systems, but they often involve unknown parameters that must be estimated from experimental data. Most calibration methods assume that measurement errors have constant variance — an assumption that is frequently violated in practice.
In this work (Sung et al, 2022), we develop a new calibration framework for inexact computer models under heteroscedastic measurement errors (i.e., non-constant variance). We derive asymptotic properties of the parameter estimators to quantify uncertainty and propose a goodness-of-fit test to detect heteroscedasticity.
Open-source implementation: HetCalibrate (GitHub).
Sung, C.-L., Barber, B. D., & Walker, B. J. (2022). SIAM/ASA J. Uncertainty Quantification, 10(4), 1733-1752.
Designing the Next Generation Rocket Injector with Physics-Informed Surrogates
High-fidelity simulations for advanced propulsion systems can generate hundreds of gigabytes of data, making design exploration computationally prohibitive.
In this work (Mak, Sung, et al, 2018), we develop a physics-informed surrogate model for turbulent flows in rocket swirl injectors. By embedding known fluid dynamics principles into the statistical framework, the method delivers accurate prediction and uncertainty quantification in roughly an hour — dramatically accelerating design cycles.
Beyond fast emulation, the model captures coupling between flow variables, enabling both reduced uncertainty and extraction of meaningful flow physics to guide next-generation injector development.
Mak, S., Sung, C.-L., Wang, X., Yeh, S.-T., Chang, Y.-H., Joseph, V. R., Yang, V., & Wu, C. F. J. (2018) Journal of the American Statistical Association, 113(524):1443-1456.
Grants
- NSF DMS 2338018: 2024-2029 (PI, $423,591) CAREER: Single-Fidelity vs. Multi-Fidelity Computer Experiments: Unveiling the Effectiveness of Multi-Fidelity Emulation
- NSF DMS 2113407: 2021-2024 (PI, $142,009) Collaborative Research: Efficient Bayesian Global Optimization with Applications to Deep Learning and Computer Experiments
