Author ORCID Identifier
0009-0004-5030-3250
Biosketch
M. Fathima received the B.Sc. degree from Kunthavai Naachiyaar Government Arts College for Women, Thanjavur, India, and the M.Sc. degree from SASTRA Deemed University, Kumbakonam, India. She earned her Ph.D. from SASTRA Deemed University, Thanjavur, India. Her research interests focus on fractional differential equations, neural network-based numerical methods, and approximation techniques.
Date of Award
14-8-2026
Document Type
Thesis
School
School of Arts, Sciences, Humanities & Education
Programme
Ph.D.-Doctoral of Philosophy
First Advisor
Dr.S.G.Venkatesh
Keywords
Fractional Differential Equations, Extreme Learning Machine, Approximation Method, Convergence Analysis, Fractional Order Shifted Legendre Polynomials
Abstract
Fractional differential equations (FDEs) have become important mathematical tools for modeling complex physical, biological, and engineering phenomena involving memory and hereditary properties. They provide more realistic representations of processes such as anomalous diffusion, viscoelasticity, and reaction-diffusion systems. However, their nonlocal characteristics make numerical solutions computationally challenging, particularly for nonlinear and complex problems. Therefore, developing accurate, stable, and computationally efficient numerical techniques is essential.
This thesis presents a novel computational framework based on the Extreme Learning Machine (ELM) technique for solving various classes of fractional differential equations. ELM is a single-hidden-layer feedforward neural network known for its fast learning capability and strong generalization performance. Unlike conventional neural network approaches, the proposed method eliminates iterative training by fixing the hidden-layer weights and biases at 1 and 0, respectively, thereby reducing computational complexity and execution time.
A major contribution of this research is the systematic development of specialized activation functions using fractional-order shifted Legendre polynomials. These functions are designed for fractional ordinary differential equations (FODEs) and systems of FODEs. For time-fractional and space-fractional partial differential equations (FPDEs), hybrid activation functions combining fractional-order shifted Legendre polynomials with classical integer-order polynomial functions are introduced. Furthermore, a two-dimensional formulation of fractional-order shifted Legendre polynomials is developed for space-time fractional partial differential equations. The Kronecker product formulation improves matrix sparsity, reduces storage requirements, and enhances computational efficiency.
In the proposed framework, the neural network output is considered a trial solution, and its fractional derivatives are substituted into the governing equations along with the corresponding initial and boundary conditions. This procedure transforms the original fractional differential equations into systems of algebraic equations. The resulting linear systems are solved using the Moore–Penrose inverse, while nonlinear problems are addressed using a nonlinear least-squares perturbation algorithm.
The theoretical properties of the considered fractional differential equations are also investigated. Existence and uniqueness of solutions are established by reformulating the governing problems into equivalent integral equations and applying fixed-point theorems. In addition, the convergence and stability of the proposed numerical framework are systematically analyzed to establish its mathematical reliability and effectiveness.
Finally, the proposed methodology is validated through several numerical experiments involving linear and nonlinear fractional differential equations. Its performance is compared with established numerical techniques, including the Legendre Neural Network Extreme Learning Machine, Müntz–Legendre polynomials, Taylor series expansion, variational iteration, Adomian decomposition, and asymptotic homotopy perturbation transform methods. The influence of the number of neurons on solution accuracy is also examined. The numerical results demonstrate that the proposed approach achieves high accuracy, improved stability, reduced computational time, and lower storage complexity, establishing it as an efficient and reliable computational technique for solving various classes of fractional differential equations.
Recommended Citation
M, Fathima Ms, "Development of An Extreme Learning Machine Technique For a Class of Fractional Differential Equations in Caputo Sense" (2026). Theses and Dissertations. 241.
https://knowledgeconnect.sastra.edu/theses/241