This work utilizes output dimensionality reduction along with information from multiple models of varying fidelities and cost to develop accurate projection-enabled multifidelity neural networks (MF-NNs) with a limited computational budget. Three approaches for MF-NNs leveraging proper orthogonal decomposition based projections are introduced: pre-training, additive, and multi-step. The MF-NN is applied to approximate the optimal design of 2D aerodynamic airfoils given performance and design requirements, demonstrating significant improvements over single-fidelity neural networks in data-scarce regimes.
@inproceedings{sella2025improving,title={Improving Neural Network Efficiency With Multifidelity and Dimensionality Reduction Techniques},author={Sella, Vignesh and O'Leary-Roseberry, Thomas and Du, Xiaosong and Guo, Mengwu and Martins, Joaquim and Ghattas, Omar and Willcox, Karen and Chaudhuri, Anirban},booktitle={AIAA SciTech Forum},year={2025},address={Orlando, FL},doi={10.2514/6.2025-2807},}
MLCSE
Projection-based multifidelity linear regression for data-scarce applications
Vignesh Sella, Julie Pham, Karen Willcox, and 1 more author
Machine Learning for Computational Science and Engineering, 2025
This work develops projection-based multifidelity linear regression methods for data-scarce applications with high-dimensional outputs. By combining high-fidelity and low-fidelity model evaluations with proper orthogonal decomposition projections, the approach achieves improved predictive accuracy over single-fidelity methods in low-data regimes. Theoretical and numerical results demonstrate robustness and accuracy improvements for surrogate modeling in computational engineering applications.
@article{sella2025projection,title={Projection-based multifidelity linear regression for data-scarce applications},author={Sella, Vignesh and Pham, Julie and Willcox, Karen and Chaudhuri, Anirban},journal={Machine Learning for Computational Science and Engineering},volume={1},number={2},pages={47},year={2025},publisher={Springer},doi={10.1007/s44379-025-00049-5},}
2024
FDS
Multifidelity linear regression for scientific machine learning from scarce data
Elizabeth Qian, Dayoung Kang, Vignesh Sella, and 1 more author
We propose a new multifidelity training approach for scientific machine learning that exploits scientific contexts where data of varying fidelities and costs are available. We use multifidelity data to define new multifidelity Monte Carlo estimators for the unknown parameters of linear regression models, providing theoretical analyses that guarantee accuracy and improved robustness to small training budgets. Numerical results demonstrate that multifidelity learned models achieve order-of-magnitude lower model variance than standard models trained on only high-fidelity data of comparable cost.
@article{qian2024multifidelity,title={Multifidelity linear regression for scientific machine learning from scarce data},author={Qian, Elizabeth and Kang, Dayoung and Sella, Vignesh and Chaudhuri, Anirban},journal={Foundations of Data Science},publisher={American Institute of Mathematical Sciences},year={2024},doi={10.3934/fods.2024049},}
2023
AIAA
Projection-based multifidelity linear regression for data-poor applications
Vignesh Sella, Julie Pham, Anirban Chaudhuri, and 1 more author
Surrogate modeling for systems with high-dimensional quantities of interest remains a challenge in situations where training data are expensive to acquire. This work develops multifidelity approaches for multivariate multi-output linear regression for data-poor applications with high-dimensional outputs, combining information from many low-cost low-fidelity model evaluations with limited expensive high-fidelity evaluations. Three projection-based multifidelity linear regression methods are implemented and contrasted, applied to approximate the surface pressure field on a hypersonic vehicle. The multifidelity approach outperforms single-fidelity regression in the low data regime with 3–10 high-fidelity samples.
@inproceedings{sella2023projection,title={Projection-based multifidelity linear regression for data-poor applications},author={Sella, Vignesh and Pham, Julie and Chaudhuri, Anirban and Willcox, Karen E},booktitle={AIAA SciTech 2023 Forum},pages={0916},year={2023},doi={10.2514/6.2023-0916},}
2021
ISC
Turbomachinery blade surrogate modeling using deep learning
Shirui Luo, Jiahuan Cui, Vignesh Sella, and 3 more authors
In International Conference on High Performance Computing, 2021
Recent work has shown that deep learning provides an alternative solution as an efficient function approximation technique for airfoil surrogate modeling. We present the feasibility of convolutional neural network (CNN) techniques for aerodynamic performance evaluation in turbo-machinery blade design. The CNN approach enables designers to fully utilize computers and statistics to interrogate and interpolate the nonlinear relationship between shapes and flow quantities, and rapidly perform optimization of the wide design space. The proposed CNN method automatically detects essential features and effectively estimates pressure loss and deviation much faster than CFD solvers.
@inproceedings{sella2021turbomachinery,title={Turbomachinery blade surrogate modeling using deep learning},author={Luo, Shirui and Cui, Jiahuan and Sella, Vignesh and Liu, Jian and Koric, Seid and Kindratenko, Volodymyr},booktitle={International Conference on High Performance Computing},pages={92--104},year={2021},organization={Springer},doi={10.1007/978-3-030-90539-2_6},}
2020
AIAA
Development of a nytrox-paraffin hybrid rocket engine
Vignesh Sella, Andrew Larkey, Abhiraj Majumder, and 7 more authors
A hybrid rocket engine designed using a nitrous oxide-oxygen mixture (Nytrox) and paraffin was developed for a high-powered rocket by the Illinois Space Society at the University of Illinois at Urbana-Champaign. The design objective is to provide 4000 N of thrust to propel a 4 kg payload to 10,000 ft for the Intercollegiate Rocketry and Engineering Competition. This work presents the design, analysis, and development of the Nytrox-paraffin hybrid rocket engine, providing students firsthand experience developing a novel oxidizer-fuel combination.
@inproceedings{sella2020nytrox,title={Development of a nytrox-paraffin hybrid rocket engine},author={Sella, Vignesh and Larkey, Andrew and Majumder, Abhiraj and Rao, Avinash and Abidi, Zavar and Rasmont, Nicolas and Randeo, Aasheesh and Liu, Miron and Moore, Avery and Lembeck, Michael F},booktitle={AIAA Propulsion and Energy 2020 Forum},pages={3729},year={2020},doi={10.2514/6.2020-3729},}
