Kensuke Kageyama

Kensuke Kageyama

Auto-generated profile, based on publicly available information

Kensuke Kageyama


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
5 May 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
THThonn Homsnit
Thonn Homsnit
KKKensuke Kageyama
Kensuke Kageyama
TKTomohisa Kojima
Tomohisa Kojima
Physics-informed neural networks have significant potential in computational mechanics, yet they remain unreliable for solving the fourth-order differential equations characteristic of thin beams. This study adopts the Euler–Bernoulli beam as a canonical diagnostic benchmark using consistent network hyperparameters to isolate the mathematical root causes of numerical instabilities and poor approximation quality across the strong-form physics-informed neural networks, the deep energy method, and the variational physics-informed neural networks. We quantify the spectral pathology for the standard strong-form physics-informed neural networks, revealing severe ill-conditioning caused by competing penalty-based boundary constraints. Meanwhile, the deep energy method inherently suffers from Hessian indefiniteness, creating saddle points that stall optimization unless mitigated by ad hoc loss scaling. For variational physics-informed neural networks, we find that the method is invariant to loss scaling only when the derivative order of boundary terms does not exceed that of domain terms. The implementation of a configuration that synergizes the strong form with elementary distance functions and a combination of first- and second-order optimizer can be established as a practical guideline for numerical robustness in thin beam problems. This approach achieves strain field accuracy comparable to standard finite element benchmarks without the need for ad hoc loss scaling. These findings offer a robust framework for the development of robust physics-based surrogate models for thin-beam applications.