Study of machine learning for analysis of heat transfer applications

Abstract

Machine learning is the outcome of the algorithmic form of statistics and a few other mathematics classes with some set of instructions, i.e., programs. This thesis explores the machine learning techniques used in the heat transfer research domain. Among the many available machine learning methods, the present work emphasizes a technique known as a Physics-informed neural network (PINN), which combines the physics associated with the problem (heat transfer) and the computational ability of a neural network. In this neural network, the loss function takes the form of a summation of different loss functions, which includes the loss in data due to PDE known as 𝐿𝑜𝑠𝑠𝑃𝐷𝐸 , loss due to boundary condition known as 𝐿𝑜𝑠𝑠𝐵𝐶 , and loss due to initial condition known as 𝐿𝑜𝑠𝑠𝐼𝐶 . In this work, the PINN technique is used to study heat transfer behavior through a slab and a fin under specified boundary and initial conditions, two of the most common problems in the thermal domain. Solutions obtained from this method are validated using analytical solutions.