Jiahao Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Jiahao Zhang has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
- Guang Lin
School of Mechanical Engineering, Purdue University, West Lafayette, USA
12 shared publications - Bongseok Kim
School of Mechanical Engineering, Purdue University, West Lafayette, Indiana 47907, United States
4 shared publications - Shiheng Zhang
Department of Applied Mathematics, University of Washington
2 shared publications Binghang Lu
1 shared publicationBong Seok Kim
1 shared publicationChristian Moya
1 shared publication
Latest Publications
- 22 October 2026
- Published by Elsevier in Computer Methods in Applied Mechanics and Engineering
- Vol. 464
Low-rank evolutionary deep neural networks via adaptive tangent-space reduction
- 22 October 2026
- Published by Elsevier in Computer Methods in Applied Mechanics and Engineering
- Vol. 464
Low-rank evolutionary deep neural networks via adaptive tangent-space reduction
- Open Access
- 7 July 2026
- Published by Elsevier in SSRN Electronic Journal
A Hyperbolic Neural Closure for M1 Radiation Transfer
In radiation transfer simulations, an M1 method achieves substantial computationalsavings by replacing the full angular transport equation with a low-order momentsystem. Because this reduced system is not closed, a closure model is requiredto represent the unknown higher-order moments using lower-order moments. Whilemachine learning (ML)-based closures can improve accuracy beyond classical analyticclosures, unconstrained learned closures may produce non-real characteristic speedsand consequently cause numerical solver breakdown. To guarantee real eigenvaluesof the Jacobian associated with ML closures, we propose a hyperbolic neural closurefor the M1 radiative transfer system. Rather than directly predicting closure terms,we parameterize the Jacobian through two neural networks: (i) a symmetric matrixnetwork and (ii) a strictly convex entropy network whose Hessian defines a positivedefinite symmetrizer. These components are combined to yield a Jacobian that issimilar to a symmetric matrix, thereby ensuring real eigenvalues. The closure isthen reconstructed by numerical integration of the learned Jacobian field along aprescribed integration path. Numerical experiments show that the proposed closurenot only achieves higher closure accuracy than classical analytic closures, but alsoimproves solution accuracy and remains stable in discontinuous Galerkin simulationsfor radiative transfer problems.
- 7 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Hyperbolic Neural Closure for M1 Radiation Transfer
In radiation transfer simulations, an M1 method achieves substantial computationalsavings by replacing the full angular transport equation with a low-order momentsystem. Because this reduced system is not closed, a closure model is requiredto represent the unknown higher-order moments using lower-order moments. Whilemachine learning (ML)-based closures can improve accuracy beyond classical analyticclosures, unconstrained learned closures may produce non-real characteristic speedsand consequently cause numerical solver breakdown. To guarantee real eigenvaluesof the Jacobian associated with ML closures, we propose a hyperbolic neural closurefor the M1 radiative transfer system. Rather than directly predicting closure terms,we parameterize the Jacobian through two neural networks: (i) a symmetric matrixnetwork and (ii) a strictly convex entropy network whose Hessian defines a positivedefinite symmetrizer. These components are combined to yield a Jacobian that issimilar to a symmetric matrix, thereby ensuring real eigenvalues. The closure isthen reconstructed by numerical integration of the learned Jacobian field along aprescribed integration path. Numerical experiments show that the proposed closurenot only achieves higher closure accuracy than classical analytic closures, but alsoimproves solution accuracy and remains stable in discontinuous Galerkin simulationsfor radiative transfer problems.
- Open Access
- 12 April 2026
- Published by Elsevier in SSRN Electronic Journal
BEKAN: Boundary condition–guaranteed evolutionary Kolmogorov–Arnold networks with radial basis functions for solving PDE problems
Deep learning has gained attention for solving partial differential equations (PDEs), but the limited interpretability of neural networks hinders precise enforcement of boundary conditions. To address this challenge, we propose BEKAN, a boundary condition-guaranteed evolutionary Kolmogorov-Arnold network with radial basis functions. BEKAN introduces three distinct and combinable approaches for rigorously incorporating Dirichlet, periodic, and Neumann boundary conditions.For Dirichlet problems, we construct boundary-embedded univariate basis functions using smooth Gaussian RBFs, encoding constraints directly at the activation level. For periodic problems, we design a sinusoidal periodic layer that exactly enforces periodicity. For Neumann problems, we formulate a least-squares evolutionary update that drives network parameters toward satisfying derivative constraints on boundaries. BEKAN also extends to general mixed boundary conditions, since the enforcement mechanisms for the three boundary types operate independently and can be assigned to distinct boundary segments. By integrating boundary-embedded RBFs, periodic layers, and evolutionary dynamics, BEKAN achieves accurate PDE simulations with rigorous boundary enforcement. Extensive numerical experiments on Dirichlet, Neumann, periodic, and mixed boundary value problems demonstrate that BEKAN consistently outperforms multilayer perceptron (MLP) and B-spline KAN baselines in both accuracy and boundary-condition satisfaction. This framework extends KAN capabilities for scientific machine learning, offering a robust and interpretable approach for solving complex PDEs under strict physical constraints.
- 12 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessBEKAN: Boundary condition–guaranteed evolutionary Kolmogorov–Arnold networks with radial basis functions for solving PDE problems
Deep learning has gained attention for solving partial differential equations (PDEs), but the limited interpretability of neural networks hinders precise enforcement of boundary conditions. To address this challenge, we propose BEKAN, a boundary condition-guaranteed evolutionary Kolmogorov-Arnold network with radial basis functions. BEKAN introduces three distinct and combinable approaches for rigorously incorporating Dirichlet, periodic, and Neumann boundary conditions.For Dirichlet problems, we construct boundary-embedded univariate basis functions using smooth Gaussian RBFs, encoding constraints directly at the activation level. For periodic problems, we design a sinusoidal periodic layer that exactly enforces periodicity. For Neumann problems, we formulate a least-squares evolutionary update that drives network parameters toward satisfying derivative constraints on boundaries. BEKAN also extends to general mixed boundary conditions, since the enforcement mechanisms for the three boundary types operate independently and can be assigned to distinct boundary segments. By integrating boundary-embedded RBFs, periodic layers, and evolutionary dynamics, BEKAN achieves accurate PDE simulations with rigorous boundary enforcement. Extensive numerical experiments on Dirichlet, Neumann, periodic, and mixed boundary value problems demonstrate that BEKAN consistently outperforms multilayer perceptron (MLP) and B-spline KAN baselines in both accuracy and boundary-condition satisfaction. This framework extends KAN capabilities for scientific machine learning, offering a robust and interpretable approach for solving complex PDEs under strict physical constraints.
- Open Access
- 7 April 2026
- Published by Elsevier in SSRN Electronic Journal
Low-Rank Evolutionary Deep Neural Networks via Adaptive Tangent-Space Reduction
Evolutionary deep neural networks (EDNNs) solve time-dependent partial differential equations by evolving the neural-network parameters sequentially in time through a local least-squares problem. Their main computational bottleneck is that each time step requires the solution of a dense linear system whose dimension equals the total number of trainable parameters. We propose a low-rank evolutionary deep neural network (LR-EDNN) method that reduces this cost through adaptive tangent-space projection. At each time step, each layer weight matrix is approximated by a rank-\(r\) truncated singular value decomposition, and the EDNN velocity is restricted to the tangent space of the corresponding fixed-rank matrix manifold. This construction replaces direct bilinear low-rank factor evolution by a linear reduced problem while preserving the sequential-in-time structure of EDNN. We show that the reduced update is the orthogonal projection of the full EDNN update in a Jacobian-induced seminorm, derive a residual decomposition identity, prove uniqueness of the induced discrete solution derivative modulo the Jacobian nullspace, and establish monotonicity with respect to the rank together with recovery of full EDNN in the full-rank limit. We further show that the one-step discrepancy between reduced and full EDNN solution updates is controlled by the tangent-space projection error up to higher-order terms. Numerical experiments on a porous-medium equation with drift, one- and two-dimensional Allen--Cahn equations, and two-dimensional viscous Burgers' equations demonstrate that LR-EDNN substantially reduces computational cost while maintaining the accuracy and qualitative fidelity of the full EDNN solver when the rank is chosen adequately.- 7 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessLow-Rank Evolutionary Deep Neural Networks via Adaptive Tangent-Space Reduction
Evolutionary deep neural networks (EDNNs) solve time-dependent partial differential equations by evolving the neural-network parameters sequentially in time through a local least-squares problem. Their main computational bottleneck is that each time step requires the solution of a dense linear system whose dimension equals the total number of trainable parameters. We propose a low-rank evolutionary deep neural network (LR-EDNN) method that reduces this cost through adaptive tangent-space projection. At each time step, each layer weight matrix is approximated by a rank-\(r\) truncated singular value decomposition, and the EDNN velocity is restricted to the tangent space of the corresponding fixed-rank matrix manifold. This construction replaces direct bilinear low-rank factor evolution by a linear reduced problem while preserving the sequential-in-time structure of EDNN. We show that the reduced update is the orthogonal projection of the full EDNN update in a Jacobian-induced seminorm, derive a residual decomposition identity, prove uniqueness of the induced discrete solution derivative modulo the Jacobian nullspace, and establish monotonicity with respect to the rank together with recovery of full EDNN in the full-rank limit. We further show that the one-step discrepancy between reduced and full EDNN solution updates is controlled by the tangent-space projection error up to higher-order terms. Numerical experiments on a porous-medium equation with drift, one- and two-dimensional Allen--Cahn equations, and two-dimensional viscous Burgers' equations demonstrate that LR-EDNN substantially reduces computational cost while maintaining the accuracy and qualitative fidelity of the full EDNN solver when the rank is chosen adequately. - Open Access
- 20 March 2026
- Published by Elsevier in SSRN Electronic Journal
Muon with Spectral Guidance: Efficient Optimization for Scientific Machine Learning
Physics-informed neural networks and neural operators often suffer from severe optimization difficulties caused by ill-conditioned gradients, multi-scale spectral behavior, and stiffness induced by physical constraints. Recently, the Muon optimizer has shown promise by performing orthogonalized updates in the singular-vector basis of the gradient, thereby improving geometric conditioning. However, its unit–singular-value updates may lead to overly aggressive steps and lack explicit stability guarantees when applied to physicsinformed learning. In this work, we propose SpecMuon, a spectral-aware optimizer that integrates Muon’s orthogonalized geometry with a mode-wise relaxed scalar auxiliary variable (RSAV) mechanism. By decomposing matrix-valued gradients into singular modes and applying RSAV updates individually along dominant spectral directions, SpecMuon adaptively regulates step sizes according to the global loss energy while preserving Muon’s scalebalancing properties. This formulation interprets optimization as a multi-mode gradient flow and enables principled control of stiff spectral components. We establish rigorous theoretical properties of SpecMuon, including a modified energy dissipation law, positivity and boundedness of auxiliary variables, and global convergence with a linear rate under the Polyak–Łojasiewicz condition. Numerical experiments on physics-informed neural networks, DeepONets, and fractional PINN–DeepONets demonstrate that SpecMuon achieves faster convergence and improved stability compared with Adam, AdamW, and the original Muon optimizer on benchmark problems such as the one-dimensional Burgers’ equation and fractional partial differential equations.
- 20 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMuon with Spectral Guidance: Efficient Optimization for Scientific Machine Learning
Physics-informed neural networks and neural operators often suffer from severe optimization difficulties caused by ill-conditioned gradients, multi-scale spectral behavior, and stiffness induced by physical constraints. Recently, the Muon optimizer has shown promise by performing orthogonalized updates in the singular-vector basis of the gradient, thereby improving geometric conditioning. However, its unit–singular-value updates may lead to overly aggressive steps and lack explicit stability guarantees when applied to physicsinformed learning. In this work, we propose SpecMuon, a spectral-aware optimizer that integrates Muon’s orthogonalized geometry with a mode-wise relaxed scalar auxiliary variable (RSAV) mechanism. By decomposing matrix-valued gradients into singular modes and applying RSAV updates individually along dominant spectral directions, SpecMuon adaptively regulates step sizes according to the global loss energy while preserving Muon’s scalebalancing properties. This formulation interprets optimization as a multi-mode gradient flow and enables principled control of stiff spectral components. We establish rigorous theoretical properties of SpecMuon, including a modified energy dissipation law, positivity and boundedness of auxiliary variables, and global convergence with a linear rate under the Polyak–Łojasiewicz condition. Numerical experiments on physics-informed neural networks, DeepONets, and fractional PINN–DeepONets demonstrate that SpecMuon achieves faster convergence and improved stability compared with Adam, AdamW, and the original Muon optimizer on benchmark problems such as the one-dimensional Burgers’ equation and fractional partial differential equations.
- Open Access
- 20 February 2026
- Published by Elsevier in SSRN Electronic Journal
Weak-Form Evolutionary Kolmogorov–Arnold Networks for Solving Partial Differential Equations
Partial differential equations (PDEs) form a central component of scientific computing.Among recent advances in deep learning, evolutionary neural networks have been developed to successively capture the temporal dynamics of time-dependent PDEs via parameter evolution. The parameter updates are obtained by solving a linear system derived from the governing equation residuals at each time step.However, strong-form evolutionary approaches can yield ill-conditioned linear systems due to pointwise residual discretization, and their computational cost scales unfavorably with the number of training samples.To address these limitations, we propose a weak-form evolutionary Kolmogorov–Arnold Network (KAN) for the scalable and accurate prediction of PDE solutions.We decouple the linear system size from the number of training samples through the weak formulation, leading to improved scalability compared to strong-form approaches.We also rigorously enforce boundary conditions by constructing the trial space with boundary-constrained KANs to satisfy Dirichlet and periodic conditions, and by incorporating derivative boundary conditions directly into the weak formulation for Neumann conditions.In conclusion, the proposed weak-form evolutionary KAN framework provides a stable and scalable approach for PDEs and contributes to scientific machine learning with potential relevance to future engineering applications.
- 20 February 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessWeak-Form Evolutionary Kolmogorov–Arnold Networks for Solving Partial Differential Equations
Partial differential equations (PDEs) form a central component of scientific computing.Among recent advances in deep learning, evolutionary neural networks have been developed to successively capture the temporal dynamics of time-dependent PDEs via parameter evolution. The parameter updates are obtained by solving a linear system derived from the governing equation residuals at each time step.However, strong-form evolutionary approaches can yield ill-conditioned linear systems due to pointwise residual discretization, and their computational cost scales unfavorably with the number of training samples.To address these limitations, we propose a weak-form evolutionary Kolmogorov–Arnold Network (KAN) for the scalable and accurate prediction of PDE solutions.We decouple the linear system size from the number of training samples through the weak formulation, leading to improved scalability compared to strong-form approaches.We also rigorously enforce boundary conditions by constructing the trial space with boundary-constrained KANs to satisfy Dirichlet and periodic conditions, and by incorporating derivative boundary conditions directly into the weak formulation for Neumann conditions.In conclusion, the proposed weak-form evolutionary KAN framework provides a stable and scalable approach for PDEs and contributes to scientific machine learning with potential relevance to future engineering applications.
- Open Access
- 9 February 2026
- Published by Elsevier in SSRN Electronic Journal
BEKAN: Boundary condition–guaranteed evolutionary Kolmogorov–Arnold networks with radial basis functions for solving PDE problems
Deep learning has gained attention for solving partial differential equations (PDEs), but the limited interpretability of neural networks hinders precise enforcement of boundary conditions.To address this challenge, we propose BEKAN, a boundary condition-guaranteed evolutionary Kolmogorov-Arnold network with radial basis functions.BEKAN introduces three distinct and combinable approaches for rigorously incorporating Dirichlet, periodic, and Neumann boundary conditions.For Dirichlet problems, we construct boundary-embedded univariate basis functions using smooth Gaussian RBFs, encoding constraints directly at the activation level.For periodic problems, we design a sinusoidal periodic layer that exactly enforces periodicity.For Neumann problems, we formulate a least-squares evolutionary update that drives network parameters toward satisfying derivative constraints on boundaries.BEKAN also extends to general mixed boundary conditions, since the enforcement mechanisms for the three boundary types operate independently and can be assigned to distinct boundary segments.By integrating boundary-embedded RBFs, periodic layers, and evolutionary dynamics, BEKAN achieves accurate PDE simulations with rigorous boundary enforcement. Extensive numerical experiments on Dirichlet, Neumann, periodic, and mixed boundary value problems demonstrate that BEKAN consistently outperforms multilayer perceptron (MLP) and B-spline KAN baselines in both accuracy and boundary-condition satisfaction.This framework extends KAN capabilities for scientific machine learning, offering a robust and interpretable approach for solving complex PDEs under strict physical constraints.
- 9 February 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessBEKAN: Boundary condition–guaranteed evolutionary Kolmogorov–Arnold networks with radial basis functions for solving PDE problems
Deep learning has gained attention for solving partial differential equations (PDEs), but the limited interpretability of neural networks hinders precise enforcement of boundary conditions.To address this challenge, we propose BEKAN, a boundary condition-guaranteed evolutionary Kolmogorov-Arnold network with radial basis functions.BEKAN introduces three distinct and combinable approaches for rigorously incorporating Dirichlet, periodic, and Neumann boundary conditions.For Dirichlet problems, we construct boundary-embedded univariate basis functions using smooth Gaussian RBFs, encoding constraints directly at the activation level.For periodic problems, we design a sinusoidal periodic layer that exactly enforces periodicity.For Neumann problems, we formulate a least-squares evolutionary update that drives network parameters toward satisfying derivative constraints on boundaries.BEKAN also extends to general mixed boundary conditions, since the enforcement mechanisms for the three boundary types operate independently and can be assigned to distinct boundary segments.By integrating boundary-embedded RBFs, periodic layers, and evolutionary dynamics, BEKAN achieves accurate PDE simulations with rigorous boundary enforcement. Extensive numerical experiments on Dirichlet, Neumann, periodic, and mixed boundary value problems demonstrate that BEKAN consistently outperforms multilayer perceptron (MLP) and B-spline KAN baselines in both accuracy and boundary-condition satisfaction.This framework extends KAN capabilities for scientific machine learning, offering a robust and interpretable approach for solving complex PDEs under strict physical constraints.
- Open Access
- 9 February 2026
- Published by Elsevier in SSRN Electronic Journal
Low-Rank Adaptation of Evolutionary Deep Neural Networks for Scalable Learning of Time-Dependent PDE Systems
A B S T R A C TEvolutionary Deep Neural Networks (EDNNs) recast time-dependent partial differential equations (PDEs) as ordinary differential equations governing neural network parameters, enabling stable long-time integration of complex, nonlinear systems. However, evolving the full parameter space at each time step requires solving a dense linear system whose size grows prohibitively with network depth and resolution, severely limiting scalability. We introduce the Low-Rank Evolutionary Deep Neural Network (LR-EDNN), which circumvents this bottleneck by constraining parameter velocities to a low-rank subspace adaptively constructed via singular value decomposition (SVD) of the current network weights. This transformation converts the computationally intractable bilinear optimization into a better conditioned, low-dimensional linear least-squares problem. Unlike static model adaptation techniques, our approach exploits the temporal smoothness of PDE solutions: optimal updates align with the dominant modes already present in the evolving weights, not with externally prescribed bases. Across benchmark problems—including the porous medium equation, Allen–Cahn phase-field dynamics, and twodimensional viscous Burgers’ flow—LR-EDNN achieves comparable accuracy to full EDNN while reducing trainable parameters by 60%-90% and delivering substantial computational speedups. These results establish low-rank parameter velocity constraints as a practical, physics-informed strategy for scalable scientific machine learning in time-dependent PDE systems.
- 9 February 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessLow-Rank Adaptation of Evolutionary Deep Neural Networks for Scalable Learning of Time-Dependent PDE Systems
A B S T R A C TEvolutionary Deep Neural Networks (EDNNs) recast time-dependent partial differential equations (PDEs) as ordinary differential equations governing neural network parameters, enabling stable long-time integration of complex, nonlinear systems. However, evolving the full parameter space at each time step requires solving a dense linear system whose size grows prohibitively with network depth and resolution, severely limiting scalability. We introduce the Low-Rank Evolutionary Deep Neural Network (LR-EDNN), which circumvents this bottleneck by constraining parameter velocities to a low-rank subspace adaptively constructed via singular value decomposition (SVD) of the current network weights. This transformation converts the computationally intractable bilinear optimization into a better conditioned, low-dimensional linear least-squares problem. Unlike static model adaptation techniques, our approach exploits the temporal smoothness of PDE solutions: optimal updates align with the dominant modes already present in the evolving weights, not with externally prescribed bases. Across benchmark problems—including the porous medium equation, Allen–Cahn phase-field dynamics, and twodimensional viscous Burgers’ flow—LR-EDNN achieves comparable accuracy to full EDNN while reducing trainable parameters by 60%-90% and delivering substantial computational speedups. These results establish low-rank parameter velocity constraints as a practical, physics-informed strategy for scalable scientific machine learning in time-dependent PDE systems.
- Open Access
- 18 January 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (2) , 325
RMFGP: A Rotated Multi-Fidelity Gaussian Process Framework for Supervised Dimension Reduction
High-dimensional surrogate modeling with limited high-fidelity data poses a major challenge in uncertainty quantification. Classical supervised dimension reduction methods often fail in this setting due to insufficient accurate observations, while low-fidelity data are abundant but biased. In this work, we propose a Rotated Multi-Fidelity Gaussian Process (RMFGP) framework that enables reliable dimension reduction and surrogate construction under severe data scarcity. The proposed method integrates nonlinear multi-fidelity Gaussian process regression with sliced average variance estimation (SAVE) to iteratively identify informative input directions. Low-fidelity data are first used to extract coarse structural information, which is exploited to rotate the input space prior to multi-fidelity model training. Predictions generated by the trained RMFGP surrogate are then used to refine the dimension reduction, allowing accurate estimation of the central sufficient dimension reduction subspace even when high-fidelity data are scarce. A Bayesian active learning strategy based on predictive uncertainty is further incorporated to adaptively select new high-fidelity samples. Numerical examples, including stochastic partial differential equations, demonstrate that RMFGP significantly improves prediction accuracy, convergence, and uncertainty propagation compared to existing Gaussian process-based dimension reduction approaches, while requiring substantially fewer high-fidelity evaluations.
- 18 January 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (2) , 325
Open AccessRMFGP: A Rotated Multi-Fidelity Gaussian Process Framework for Supervised Dimension Reduction
High-dimensional surrogate modeling with limited high-fidelity data poses a major challenge in uncertainty quantification. Classical supervised dimension reduction methods often fail in this setting due to insufficient accurate observations, while low-fidelity data are abundant but biased. In this work, we propose a Rotated Multi-Fidelity Gaussian Process (RMFGP) framework that enables reliable dimension reduction and surrogate construction under severe data scarcity. The proposed method integrates nonlinear multi-fidelity Gaussian process regression with sliced average variance estimation (SAVE) to iteratively identify informative input directions. Low-fidelity data are first used to extract coarse structural information, which is exploited to rotate the input space prior to multi-fidelity model training. Predictions generated by the trained RMFGP surrogate are then used to refine the dimension reduction, allowing accurate estimation of the central sufficient dimension reduction subspace even when high-fidelity data are scarce. A Bayesian active learning strategy based on predictive uncertainty is further incorporated to adaptively select new high-fidelity samples. Numerical examples, including stochastic partial differential equations, demonstrate that RMFGP significantly improves prediction accuracy, convergence, and uncertainty propagation compared to existing Gaussian process-based dimension reduction approaches, while requiring substantially fewer high-fidelity evaluations.