Five students have been selected from among the applicants to the EASIAM Student Paper Prize and will present their work in this session. Based on these presentations, three final prize recipients will be selected.
Finalist Presentations
-
Computer-Assisted Proof of the Simplicity of the Second Dirichlet Eigenvalue for Non-Equilateral Triangles
Ryoki Endo (Niigata University)
This study proves that the second eigenvalue of the Dirichlet Laplacian is simple for every non-equilateral triangle, resolving a conjecture of R. Laugesen and B. Siudeja. The main difficulties arise for nearly equilateral triangles, where the second and third eigenvalues form a tight cluster, and for collapsing triangles, where the eigenvalues diverge. We first introduce a difference-quotient formula based on the Hadamard shape derivative to handle the clustered eigenvalues of nearly equilateral triangles. For collapsing triangles, we derive rigorous eigenvalue bounds using one-dimensional Schrödinger operators. Combining these estimates with verified numerical computation establishes the desired simplicity result for all non-equilateral triangles. This is joint work with Xuefeng Liu (Tokyo Woman’s Christian University, Japan).
-
CEENs: Causality-Enforced Evolutional Networks for Solving Time-Dependent Partial Differential Equations
Heechang Kim (Pohang University of Science and Technology)
Physics-informed neural networks (PINNs) have shown promise for solving partial differential equations, but they often struggle with long-time integration because their conventional training procedure does not explicitly respect temporal causality. In this talk, we present causality-enforced evolutional networks (CEENs), which represent the solution as a sequence of neural states and train them progressively from the initial condition forward in time. By constructing local loss functions from the integral form of the governing equation, CEENs prevent later-time states from being learned before earlier states have been accurately resolved. Numerical experiments on several time-dependent PDEs demonstrate that CEENs substantially improve long-time accuracy while reducing computational cost and memory usage compared with conventional PINNs. We also discuss a causality-preserving parallelization strategy and the error behavior of CEENs with respect to the time-step size and training tolerance.
-
DRM Revisited: A Complete Error Analysis
Peiying Wu (Wuhan University)
It is widely known that the error analysis for deep learning involves approximation, statistical, and optimization errors. However, it is challenging to combine them together due to overparameterization. In this paper, we address this gap by providing a comprehensive error analysis of the Deep Ritz Method (DRM). Specifically, we investigate a foundational question in the theoretical analysis of DRM under the overparameterized regime: given a target precision level, how can one determine the appropriate number of training samples, the key architectural parameters of the neural networks, the step size for the projected gradient descent optimization procedure, and the requisite number of iterations, such that the output of the gradient descent process closely approximates the true solution of the underlying partial differential equation to the specified precision?
-
Parallelization of the Random Feature Method with Engineering Applications
Yifei Sun (Soochow University)
The Random Feature Method (RFM) has emerged as an efficient mesh-free numerical approach for solving partial differential equations, offering strong flexibility in handling complex geometries. However, its application to large-scale engineering problems is often constrained by the substantial computational and memory costs associated with large least-squares systems, making efficient parallelization essential. This talk presents recent advances in the parallelization of the RFM based on domain decomposition and high-performance computing techniques. The proposed framework enables distributed computation while preserving the accuracy and robustness of the original method. Building on this parallel framework, the RFM is applied to a variety of engineering problems and further integrated into industrial software platforms to support large-scale scientific computing and computer-aided engineering applications.
-
Interpolation, approximation, and controllability of deep neural networks
Jingpu Cheng (National University of Singapore)
Viewing deep residual neural networks as controlled dynamical systems, we use controllability theory to study two notions of expressive power arising from supervised learning: universal interpolation of finite data sets and universal approximation of target functions. In this framework, universal interpolation becomes a problem of simultaneously controlling an ensemble of data points. For globally Lipschitz, affine-invariant control families, we show that exact universal interpolation holds if and only if the family contains a nonlinear vector field. This applies to a broad class of ResNets with nonlinear activations. We also show that, for general control systems, universal interpolation and universal approximation do not imply one another, while their equivalence can be recovered under suitable regularity assumptions and a uniform bound on interpolation time.