Department of Computer Science and Technology

Technical reports

Neural solver for Partial Differential Equations in scientific simulation

Zheyuan Hu

November 2026, 63 pages

This technical report is based on a dissertation submitted June 2026 by the author for the degree of Master of Engineering (Computer Science Tripos) to the University of Cambridge, Magdalene College.

The author was supervised by Dr Chenliang Zhou, with partial inspiration for this work provided by Dr Fangcheng Zhong.

DOIhttps://doi.org/10.48456/tr-1012

Abstract

Partial Differential Equations offer a mathematically rigorous framework for graphical, biological, and physical phenomena. However, the numerical approaches are hindered by the restrictions of linearity, dimensionality, and domain-specific tuning. I present a generalised framework to solve coupled and nonlinear PDEs with auto tuning, alongside its application to both forward and inverse scientific simulation problems. With an empirical framework for stability proposed, I mitigate solver sensitivity to input perturbations by introducing regularisation strategies. Through a comprehensive survey of PDEs with their applications, different types of state-of-the-art baseline methods, including FDM, PINN, FNO, GNN-based solvers, and the learning theory behind them, I systematically evaluate the proposed method across multiple critical dimensions, i.e. effectiveness quantified via error metrics, efficiency characterised by computational cost or convergence rate, and stability. The results contrast each solver strengths and limitations, highlighting the high precision, flexible kernel design, and utility in various scientific simulation contexts of the proposed method.

Full text

PDF (2.7 MB)

BibTeX record

@TechReport{UCAM-CL-TR-1012,
  author =	 {Hu, Zheyuan},
  title = 	 {{Neural solver for Partial Differential Equations in
         	   scientific simulation}},
  year = 	 2026,
  month = 	 nov,
  url = 	 {https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-1012.pdf},
  institution =  {University of Cambridge, Computer Laboratory},
  doi = 	 {10.48456/tr-1012},
  number = 	 {UCAM-CL-TR-1012}
}