Zhaojun Bai
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Zhaojun Bai has a total of 86 co-authors. Here are the top co-authors by number of shared publications:
- Ren-Cang Li
Department of Mathematics, University of Texas at Arlington, Arlington, USA
6 shared publications - Dong Min Roh
Department of Mathematics University of California 95616 Davis CA USA
5 shared publications - Bradley Tom
California Department of Water Resources, 1516 9th Street, Sacramento, CA 95814, USA
4 shared publications - Jamie Anderson
California Department of Water Resources, 1516 9th Street, Sacramento, CA 95814, USA
4 shared publications - Minxue He
California Department of Water Resources, Sacramento, CA 95814, USA
4 shared publications - Peyman Namadi
California Department of Water Resources, 1516 9th Street, Sacramento, CA 95814, USA
4 shared publications
Latest Publications
- 17 February 2026
- Published by Springer Nature in BIT Numerical Mathematics
- Vol. 66 (1) , 17
Fair principal component analysis via eigenvalue optimization
- 17 February 2026
- Published by Springer Nature in BIT Numerical Mathematics
- Vol. 66 (1) , 17
Fair principal component analysis via eigenvalue optimization
- Open Access
Mixed Precision Photonic Computing with 3D Electronic-Photonic Integrated Circuits
We propose advancing photonic in-memory computing through three-dimensional photonic-electronic integrated circuits using phase-change materials (PCM) and AlGaAs-CMOS technology. These circuits offer high precision (greater than 12 bits), scalability (greater than 1024 by 1024), and massive parallelism (greater than 1 million operations) across the wavelength, spatial, and temporal domains at ultra-low power (less than 1 watt per PetaOPS). Monolithically integrated hybrid PCM-AlGaAs memory resonators handle coarse-precision iterations (greater than 5-bit most significant bit precision) through reversible PCM phase transitions. Electro-optic memristive tuning enables fine-precision updates (greater than 8-bit least significant bit precision), resulting in over 12-bit precision for in-memory computing. The use of low-loss PCM (less than 0.01 dB per cm) and electro-optical tuning yields memristive optical resonators with high Q-factors (greater than 1 million), low insertion loss, and low tuning power. A W by W photonic tensor core composed of PCM-AlGaAs memresonators performs general matrix multiplication (GEMM) across W wavelengths from optical frequency combs, with minimal crosstalk and loss. Hierarchical scaling in the wavelength domain (K) and spatial domain (L) enables this system to address high-dimensional (N) scientific partial differential equation (PDE) problems in a single constant-time operation, compared to the conventional quadratic-time (N squared) computational complexity.
Open AccessMixed Precision Photonic Computing with 3D Electronic-Photonic Integrated Circuits
We propose advancing photonic in-memory computing through three-dimensional photonic-electronic integrated circuits using phase-change materials (PCM) and AlGaAs-CMOS technology. These circuits offer high precision (greater than 12 bits), scalability (greater than 1024 by 1024), and massive parallelism (greater than 1 million operations) across the wavelength, spatial, and temporal domains at ultra-low power (less than 1 watt per PetaOPS). Monolithically integrated hybrid PCM-AlGaAs memory resonators handle coarse-precision iterations (greater than 5-bit most significant bit precision) through reversible PCM phase transitions. Electro-optic memristive tuning enables fine-precision updates (greater than 8-bit least significant bit precision), resulting in over 12-bit precision for in-memory computing. The use of low-loss PCM (less than 0.01 dB per cm) and electro-optical tuning yields memristive optical resonators with high Q-factors (greater than 1 million), low insertion loss, and low tuning power. A W by W photonic tensor core composed of PCM-AlGaAs memresonators performs general matrix multiplication (GEMM) across W wavelengths from optical frequency combs, with minimal crosstalk and loss. Hierarchical scaling in the wavelength domain (K) and spatial domain (L) enables this system to address high-dimensional (N) scientific partial differential equation (PDE) problems in a single constant-time operation, compared to the conventional quadratic-time (N squared) computational complexity.
- Open Access
EPIC: a provable accelerated Eigensolver based on Preconditioning and Implicit Convexity
This paper is concerned with the extraction of the smallest eigenvalue and the corresponding eigenvector of a symmetric positive definite matrix pencil. We reveal implicit convexity of the eigenvalue problem in Euclidean space. A provable accelerated eigensolver based on preconditioning and implicit convexity (EPIC) is proposed. Theoretical analysis shows the acceleration of EPIC with the rate of convergence resembling the expected rate of convergence of the well-known locally optimal preconditioned conjugate gradient (LOPCG). A complete proof of the expected rate of convergence of LOPCG is elusive so far. Numerical results confirm our theoretical findings of EPIC.
Open AccessEPIC: a provable accelerated Eigensolver based on Preconditioning and Implicit Convexity
This paper is concerned with the extraction of the smallest eigenvalue and the corresponding eigenvector of a symmetric positive definite matrix pencil. We reveal implicit convexity of the eigenvalue problem in Euclidean space. A provable accelerated eigensolver based on preconditioning and implicit convexity (EPIC) is proposed. Theoretical analysis shows the acceleration of EPIC with the rate of convergence resembling the expected rate of convergence of the well-known locally optimal preconditioned conjugate gradient (LOPCG). A complete proof of the expected rate of convergence of LOPCG is elusive so far. Numerical results confirm our theoretical findings of EPIC.
- Open Access
A Self-Consistent Field Solution for Robust Common Spatial Pattern Analysis
The common spatial pattern analysis (CSP) is a widely used signal processing technique in brain-computer interface (BCI) systems to increase the signal-to-noise ratio in electroencephalogram (EEG) recordings. Despite its popularity, the CSP's performance is often hindered by the nonstationarity and artifacts in EEG signals. The minmax CSP improves the robustness of the CSP by using data-driven covariance matrices to accommodate the uncertainties. We show that by utilizing the optimality conditions, the minmax CSP can be recast as an eigenvector-dependent nonlinear eigenvalue problem (NEPv). We introduce a self-consistent field (SCF) iteration with line search that solves the NEPv of the minmax CSP. Local quadratic convergence of the SCF for solving the NEPv is illustrated using synthetic datasets. More importantly, experiments with real-world EEG datasets show the improved motor imagery classification rates and shorter running time of the proposed SCF-based solver compared to the existing algorithm for the minmax CSP.
Open AccessA Self-Consistent Field Solution for Robust Common Spatial Pattern Analysis
The common spatial pattern analysis (CSP) is a widely used signal processing technique in brain-computer interface (BCI) systems to increase the signal-to-noise ratio in electroencephalogram (EEG) recordings. Despite its popularity, the CSP's performance is often hindered by the nonstationarity and artifacts in EEG signals. The minmax CSP improves the robustness of the CSP by using data-driven covariance matrices to accommodate the uncertainties. We show that by utilizing the optimality conditions, the minmax CSP can be recast as an eigenvector-dependent nonlinear eigenvalue problem (NEPv). We introduce a self-consistent field (SCF) iteration with line search that solves the NEPv of the minmax CSP. Local quadratic convergence of the SCF for solving the NEPv is illustrated using synthetic datasets. More importantly, experiments with real-world EEG datasets show the improved motor imagery classification rates and shorter running time of the proposed SCF-based solver compared to the existing algorithm for the minmax CSP.
- Open Access
A Bi-level Nonlinear Eigenvector Algorithm for Wasserstein Discriminant Analysis
Much like the classical Fisher linear discriminant analysis (LDA), the recently proposed Wasserstein discriminant analysis (WDA) is a linear dimensionality reduction method that seeks a projection matrix to maximize the dispersion of different data classes and minimize the dispersion of same data classes via a bi-level optimization. In contrast to LDA, WDA can account for both global and local interconnections between data classes by using the underlying principles of optimal transport. In this paper, a bi-level nonlinear eigenvector algorithm (WDA-nepv) is presented to fully exploit the structures of the bi-level optimization of WDA. The inner level of WDA-nepv for computing the optimal transport matrices is formulated as an eigenvector-dependent nonlinear eigenvalue problem (NEPv), and meanwhile, the outer level for trace ratio optimizations is formulated as another NEPv. Both NEPvs can be computed efficiently under the self-consistent field (SCF) framework. WDA-nepv is derivative-free and surrogate-model-free when compared with existing algorithms. Convergence analysis of the proposed WDA-nepv justifies the utilization of the SCF for solving the bi-level optimization of WDA. Numerical experiments with synthetic and real-life datasets demonstrate the classification accuracy and scalability of WDA-nepv.
Open AccessA Bi-level Nonlinear Eigenvector Algorithm for Wasserstein Discriminant Analysis
Much like the classical Fisher linear discriminant analysis (LDA), the recently proposed Wasserstein discriminant analysis (WDA) is a linear dimensionality reduction method that seeks a projection matrix to maximize the dispersion of different data classes and minimize the dispersion of same data classes via a bi-level optimization. In contrast to LDA, WDA can account for both global and local interconnections between data classes by using the underlying principles of optimal transport. In this paper, a bi-level nonlinear eigenvector algorithm (WDA-nepv) is presented to fully exploit the structures of the bi-level optimization of WDA. The inner level of WDA-nepv for computing the optimal transport matrices is formulated as an eigenvector-dependent nonlinear eigenvalue problem (NEPv), and meanwhile, the outer level for trace ratio optimizations is formulated as another NEPv. Both NEPvs can be computed efficiently under the self-consistent field (SCF) framework. WDA-nepv is derivative-free and surrogate-model-free when compared with existing algorithms. Convergence analysis of the proposed WDA-nepv justifies the utilization of the SCF for solving the bi-level optimization of WDA. Numerical experiments with synthetic and real-life datasets demonstrate the classification accuracy and scalability of WDA-nepv.
- Open Access
Salinity Modeling Using Deep Learning with Data Augmentation and Transfer Learning
Salinity management in estuarine systems is crucial for developing effective water-management strategies to maintain compliance and understand the impact of salt intrusion on water quality and availability. Understanding the temporal and spatial variations of salinity is a keystone of salinity-management practices. Process-based numerical models have been traditionally used to estimate the variations in salinity in estuarine environments. Advances in data-driven models (e.g., deep learning models) make them effective and efficient alternatives to process-based models. However, a discernible research gap exists in applying these advanced techniques to salinity modeling. The current study seeks to address this gap by exploring the innovative use of deep learning with data augmentation and transfer learning in salinity modeling, exemplified at 23 key salinity locations in the Sacramento–San Joaquin Delta which is the hub of the water-supply system of California. Historical, simulated (via a hydrodynamics and water quality model), and perturbed (to create a range of hydroclimatic and operational scenarios for data-augmentation purposes) flow, and salinity data are used to train a baseline multi-layer perceptron (MLP) and a deep learning Residual Long-Short-Term Memory (Res-LSTM) network. Four other deep learning models including LSTM, Residual Network (ResNet), Gated Recurrent Unit (GRU), and Residual GRU (Res-GRU) are also examined. Results indicate that models pre-trained using augmented data demonstrate improved performance over models trained from scratch using only historical data (e.g., median Nash–Sutcliffe efficiency increased from around 0.5 to above 0.9). Moreover, the five deep learning models further boost the salinity estimation performance in comparison with the baseline MLP model, though the performance of the latter is acceptable. The models trained using augmented data are then (a) used to develop a web-based Salinity Dashboard (Dashboard) tool that allows the users (including those with no machine learning background) to quickly screen multiple management scenarios by altering inputs and visualizing the resulting salinity simulations interactively, and (b) transferred and adapted to estimate observed salinity. The study shows that transfer learning results more accurately replicate the observations compared to their counterparts from models trained from scratch without knowledge learned and transferred from augmented data (e.g., median Nash–Sutcliffe efficiency increased from around 0.4 to above 0.9). Overall, the study illustrates that deep learning models, particularly when pre-trained using augmented data, are promising supplements to existing process-based models in estuarine salinity modeling, while the Dashboard enables user engagement with those pre-trained models to inform decision-making efficiently and effectively.
Open AccessSalinity Modeling Using Deep Learning with Data Augmentation and Transfer Learning
Salinity management in estuarine systems is crucial for developing effective water-management strategies to maintain compliance and understand the impact of salt intrusion on water quality and availability. Understanding the temporal and spatial variations of salinity is a keystone of salinity-management practices. Process-based numerical models have been traditionally used to estimate the variations in salinity in estuarine environments. Advances in data-driven models (e.g., deep learning models) make them effective and efficient alternatives to process-based models. However, a discernible research gap exists in applying these advanced techniques to salinity modeling. The current study seeks to address this gap by exploring the innovative use of deep learning with data augmentation and transfer learning in salinity modeling, exemplified at 23 key salinity locations in the Sacramento–San Joaquin Delta which is the hub of the water-supply system of California. Historical, simulated (via a hydrodynamics and water quality model), and perturbed (to create a range of hydroclimatic and operational scenarios for data-augmentation purposes) flow, and salinity data are used to train a baseline multi-layer perceptron (MLP) and a deep learning Residual Long-Short-Term Memory (Res-LSTM) network. Four other deep learning models including LSTM, Residual Network (ResNet), Gated Recurrent Unit (GRU), and Residual GRU (Res-GRU) are also examined. Results indicate that models pre-trained using augmented data demonstrate improved performance over models trained from scratch using only historical data (e.g., median Nash–Sutcliffe efficiency increased from around 0.5 to above 0.9). Moreover, the five deep learning models further boost the salinity estimation performance in comparison with the baseline MLP model, though the performance of the latter is acceptable. The models trained using augmented data are then (a) used to develop a web-based Salinity Dashboard (Dashboard) tool that allows the users (including those with no machine learning background) to quickly screen multiple management scenarios by altering inputs and visualizing the resulting salinity simulations interactively, and (b) transferred and adapted to estimate observed salinity. The study shows that transfer learning results more accurately replicate the observations compared to their counterparts from models trained from scratch without knowledge learned and transferred from augmented data (e.g., median Nash–Sutcliffe efficiency increased from around 0.4 to above 0.9). Overall, the study illustrates that deep learning models, particularly when pre-trained using augmented data, are promising supplements to existing process-based models in estuarine salinity modeling, while the Dashboard enables user engagement with those pre-trained models to inform decision-making efficiently and effectively.
- Open Access
Physics-Informed Neural Networks-Based Salinity Modeling in the Sacramento–San Joaquin Delta of California
Salinity in estuarine environments has been traditionally simulated using process-based models. More recently, data-driven models including artificial neural networks (ANNs) have been developed for simulating salinity. Compared to process-based models, ANNs yield faster salinity simulations with comparable accuracy. However, ANNs are often purely data-driven and not constrained by physical laws, making it difficult to interpret the causality between input and output data. Physics-informed neural networks (PINNs) are emerging machine-learning models to integrate the benefits of both process-based models and data-driven ANNs. PINNs can embed the knowledge of physical laws in terms of the partial differential equations (PDE) that govern the dynamics of salinity transport into the training of the neural networks. This study explores the application of PINNs in salinity modeling by incorporating the one-dimensional advection–dispersion salinity transport equation into the neural networks. Two PINN models are explored in this study, namely PINNs and FoNets. PINNs are multilayer perceptrons (MLPs) that incorporate the advection–dispersion equation, while FoNets are an extension of PINNs with an additional encoding layer. The exploration is exemplified at four study locations in the Sacramento–San Joaquin Delta of California: Pittsburg, Chipps Island, Port Chicago, and Martinez. Both PINN models and benchmark ANNs are trained and tested using simulated daily salinity from 1991 to 2015 at study locations. Results indicate that PINNs and FoNets outperform the benchmark ANNs in simulating salinity at the study locations. Specifically, PINNs and FoNets have lower absolute biases and higher correlation coefficients and Nash–Sutcliffe efficiency values than ANNs. In addition, PINN models overcome some limitations of purely data-driven ANNs (e.g., neuron saturation) and generate more realistic salinity simulations. Overall, this study demonstrates the potential of PINNs to supplement existing process-based and ANN models in providing accurate and timely salinity estimation.
Open AccessPhysics-Informed Neural Networks-Based Salinity Modeling in the Sacramento–San Joaquin Delta of California
Salinity in estuarine environments has been traditionally simulated using process-based models. More recently, data-driven models including artificial neural networks (ANNs) have been developed for simulating salinity. Compared to process-based models, ANNs yield faster salinity simulations with comparable accuracy. However, ANNs are often purely data-driven and not constrained by physical laws, making it difficult to interpret the causality between input and output data. Physics-informed neural networks (PINNs) are emerging machine-learning models to integrate the benefits of both process-based models and data-driven ANNs. PINNs can embed the knowledge of physical laws in terms of the partial differential equations (PDE) that govern the dynamics of salinity transport into the training of the neural networks. This study explores the application of PINNs in salinity modeling by incorporating the one-dimensional advection–dispersion salinity transport equation into the neural networks. Two PINN models are explored in this study, namely PINNs and FoNets. PINNs are multilayer perceptrons (MLPs) that incorporate the advection–dispersion equation, while FoNets are an extension of PINNs with an additional encoding layer. The exploration is exemplified at four study locations in the Sacramento–San Joaquin Delta of California: Pittsburg, Chipps Island, Port Chicago, and Martinez. Both PINN models and benchmark ANNs are trained and tested using simulated daily salinity from 1991 to 2015 at study locations. Results indicate that PINNs and FoNets outperform the benchmark ANNs in simulating salinity at the study locations. Specifically, PINNs and FoNets have lower absolute biases and higher correlation coefficients and Nash–Sutcliffe efficiency values than ANNs. In addition, PINN models overcome some limitations of purely data-driven ANNs (e.g., neuron saturation) and generate more realistic salinity simulations. Overall, this study demonstrates the potential of PINNs to supplement existing process-based and ANN models in providing accurate and timely salinity estimation.
- Open Access
Scalable Spectral Clustering with Group Fairness Constraints
There are synergies of research interests and industrial efforts in modeling fairness and correcting algorithmic bias in machine learning. In this paper, we present a scalable algorithm for spectral clustering (SC) with group fairness constraints. Group fairness is also known as statistical parity where in each cluster, each protected group is represented with the same proportion as in the entirety. While FairSC algorithm (Kleindessner et al., 2019) is able to find the fairer clustering, it is compromised by high costs due to the kernels of computing nullspaces and the square roots of dense matrices explicitly. We present a new formulation of underlying spectral computation by incorporating nullspace projection and Hotelling's deflation such that the resulting algorithm, called s-FairSC, only involves the sparse matrix-vector products and is able to fully exploit the sparsity of the fair SC model. The experimental results on the modified stochastic block model demonstrate that s-FairSC is comparable with FairSC in recovering fair clustering. Meanwhile, it is sped up by a factor of 12 for moderate model sizes. s-FairSC is further demonstrated to be scalable in the sense that the computational costs of s-FairSC only increase marginally compared to the SC without fairness constraints.
Open AccessScalable Spectral Clustering with Group Fairness Constraints
There are synergies of research interests and industrial efforts in modeling fairness and correcting algorithmic bias in machine learning. In this paper, we present a scalable algorithm for spectral clustering (SC) with group fairness constraints. Group fairness is also known as statistical parity where in each cluster, each protected group is represented with the same proportion as in the entirety. While FairSC algorithm (Kleindessner et al., 2019) is able to find the fairer clustering, it is compromised by high costs due to the kernels of computing nullspaces and the square roots of dense matrices explicitly. We present a new formulation of underlying spectral computation by incorporating nullspace projection and Hotelling's deflation such that the resulting algorithm, called s-FairSC, only involves the sparse matrix-vector products and is able to fully exploit the sparsity of the fair SC model. The experimental results on the modified stochastic block model demonstrate that s-FairSC is comparable with FairSC in recovering fair clustering. Meanwhile, it is sped up by a factor of 12 for moderate model sizes. s-FairSC is further demonstrated to be scalable in the sense that the computational costs of s-FairSC only increase marginally compared to the SC without fairness constraints.
- Open Access
2D Eigenvalue Problem III: Convergence Analysis of the 2D Rayleigh Quotient Iteration
In Part I of this paper, we introduced a two dimensional eigenvalue problem (2DEVP) of a matrix pair and investigated its fundamental theory such as existence, variational characterization and number of 2D-eigenvalues. In Part II, we proposed a Rayleigh quotient iteration (RQI)-like algorithm (2DRQI) for computing a 2D-eigentriplet of the 2DEVP near a prescribed point, and discussed applications of 2DEVP and 2DRQI for solving the minimax problem of Rayleigh quotients, and computing the distance to instability. In this third part, we present convergence analysis of the 2DRQI. We show that under some mild conditions, the 2DRQI is locally quadratically convergent for computing a nonsingular 2D-eigentriplet.
Open Access2D Eigenvalue Problem III: Convergence Analysis of the 2D Rayleigh Quotient Iteration
In Part I of this paper, we introduced a two dimensional eigenvalue problem (2DEVP) of a matrix pair and investigated its fundamental theory such as existence, variational characterization and number of 2D-eigenvalues. In Part II, we proposed a Rayleigh quotient iteration (RQI)-like algorithm (2DRQI) for computing a 2D-eigentriplet of the 2DEVP near a prescribed point, and discussed applications of 2DEVP and 2DRQI for solving the minimax problem of Rayleigh quotients, and computing the distance to instability. In this third part, we present convergence analysis of the 2DRQI. We show that under some mild conditions, the 2DRQI is locally quadratically convergent for computing a nonsingular 2D-eigentriplet.
- Open Access
Novel Salinity Modeling Using Deep Learning for the Sacramento–San Joaquin Delta of California
Water resources management in estuarine environments for water supply and environmental protection typically requires estimates of salinity for various flow and operational conditions. This study develops and applies two novel deep learning (DL) models, a residual long short-term memory (Res-LSTM) network, and a residual gated recurrent unit (Res-GRU) model, in estimating the spatial and temporal variations of salinity. Four other machine learning (ML) models, previously developed and reported, consisting of multi-layer perceptron (MLP), residual network (ResNet), LSTM, and GRU are utilized as the baseline models to benchmark the performance of the two novel models. All six models are applied at 23 study locations in the Sacramento–San Joaquin Delta (Delta), the hub of California’s water supply system. Model input features include observed or calculated tidal stage (water level), flow and salinity at model upstream boundaries, salinity control gate operations, crop consumptive use, and pumping for the period of 2001–2019. Meanwhile, field observations of salinity at the study locations during the same period are also utilized for the development of the predictive use of the models. Results indicate that the proposed DL models generally outperform the baseline models in simulating and predicting salinity on both daily and hourly scales at the study locations. The absolute bias is generally less than 5%. The correlation coefficients and Nash–Sutcliffe efficiency values are close to 1. Particularly, Res-LSTM has slightly superior performance over Res-GRU. Moreover, the study investigates the overfitting issues of both the DL and baseline models. The investigation indicates that overfitting is not notable. Finally, the study compares the performance of Res-LSTM against that of an operational process-based salinity model. It is shown Res-LSTM outperforms the process-based model consistently across all study locations. Overall, the study demonstrates the feasibility of DL-based models in supplementing the existing operational models in providing accurate and real-time estimates of salinity to inform water management decision making.
Open AccessNovel Salinity Modeling Using Deep Learning for the Sacramento–San Joaquin Delta of California
Water resources management in estuarine environments for water supply and environmental protection typically requires estimates of salinity for various flow and operational conditions. This study develops and applies two novel deep learning (DL) models, a residual long short-term memory (Res-LSTM) network, and a residual gated recurrent unit (Res-GRU) model, in estimating the spatial and temporal variations of salinity. Four other machine learning (ML) models, previously developed and reported, consisting of multi-layer perceptron (MLP), residual network (ResNet), LSTM, and GRU are utilized as the baseline models to benchmark the performance of the two novel models. All six models are applied at 23 study locations in the Sacramento–San Joaquin Delta (Delta), the hub of California’s water supply system. Model input features include observed or calculated tidal stage (water level), flow and salinity at model upstream boundaries, salinity control gate operations, crop consumptive use, and pumping for the period of 2001–2019. Meanwhile, field observations of salinity at the study locations during the same period are also utilized for the development of the predictive use of the models. Results indicate that the proposed DL models generally outperform the baseline models in simulating and predicting salinity on both daily and hourly scales at the study locations. The absolute bias is generally less than 5%. The correlation coefficients and Nash–Sutcliffe efficiency values are close to 1. Particularly, Res-LSTM has slightly superior performance over Res-GRU. Moreover, the study investigates the overfitting issues of both the DL and baseline models. The investigation indicates that overfitting is not notable. Finally, the study compares the performance of Res-LSTM against that of an operational process-based salinity model. It is shown Res-LSTM outperforms the process-based model consistently across all study locations. Overall, the study demonstrates the feasibility of DL-based models in supplementing the existing operational models in providing accurate and real-time estimates of salinity to inform water management decision making.