Publications

We present a novel learnable-differentiable finite volume solver that significantly accelerates computational fluid dynamics simulations. By combining traditional finite volume methods with differentiable programming and machine learning techniques, our approach enables end-to-end optimization of flow simulations while maintaining physical accuracy and computational efficiency.

We introduce SlotPi, a novel physics-informed framework that combines object-centric representation learning with physical reasoning. By integrating physics principles into slot attention mechanisms, our approach enables more interpretable and physically consistent object-centric reasoning in complex dynamic systems.

We present a novel conservation-informed graph learning approach for predicting spatiotemporal dynamics in complex systems. By incorporating physical conservation laws into graph neural networks, our method achieves superior accuracy in modeling spatiotemporal phenomena while maintaining physical consistency, advancing the field of AI4Science.

We introduce P²C Net, a novel PDE-preserved coarse correction network that efficiently predicts spatiotemporal dynamics. Our method combines coarse-scale predictions with fine-scale corrections while preserving the underlying physical constraints, achieving significant computational speedup without sacrificing accuracy in spatiotemporal modeling tasks.

We present PhyMPGN, a physics-encoded message passing graph network designed for solving spatiotemporal PDE systems. Our approach integrates physical principles directly into the message passing framework, enabling more accurate and physically consistent predictions for complex spatiotemporal dynamics while maintaining computational efficiency.

We present PDEformer, a transformer-based foundation model designed specifically for solving one-dimensional partial differential equations. By leveraging the transformer architecture's sequence modeling capabilities, our approach demonstrates strong generalization across diverse PDE families and boundary conditions, paving the way for unified neural PDE solvers.

We present a novel approach for predicting transonic flow over supercritical airfoils using geometric-encoding and deep-learning strategies. Our method combines advanced geometric encoding techniques with deep neural networks to accurately predict complex flow patterns around airfoils, enabling efficient aerodynamic analysis and design optimization for transonic flight conditions.

We present a novel deep learning-based reduced order model for three-dimensional unsteady flow simulation. Our approach combines mesh transformation techniques with stitching methods to efficiently model complex fluid dynamics, enabling fast and accurate predictions of unsteady flows while maintaining computational efficiency and physical accuracy.

We propose a novel approach for transportation origin-destination (OD) demand estimation using quasi-sparsity techniques. Our method addresses the challenge of estimating OD flows in transportation networks by leveraging advanced optimization algorithms and sparsity constraints, enabling more accurate and efficient transportation planning and traffic management.

We introduce Meta-auto-decoder, a novel meta-learning framework for solving parametric partial differential equations (PDEs). Our approach combines meta-learning with auto-decoder architectures to efficiently adapt to new PDE parameters and boundary conditions, enabling rapid solution of parametric PDE families with minimal computational overhead and superior generalization capabilities.

We propose a universal Physics-Informed Neural Networks (PINNs) method for solving partial differential equations with point sources. Our approach addresses the challenge of handling singular point sources in PDEs by developing a novel neural network architecture that can effectively capture the complex behavior around singularities while maintaining high accuracy across the entire domain.

We analyze the convergence properties of the augmented decomposition algorithm for solving large-scale optimization problems. Our theoretical analysis provides convergence guarantees and establishes convergence rates for this important class of decomposition methods, with applications to distributed optimization and ML infrastructure.