Journal article 144 views 50 downloads
Data-driven inverse modelling through neural network (deep learning) and computational heat transfer / Hamid Tamaddon-Jahromi; Neeraj Kavan Chakshu; Igor Sazonov; Llion Evans; Hywel Thomas; Perumal Nithiarasu
Computer Methods in Applied Mechanics and Engineering, Volume: 369
PDF | Version of Record
Released under the terms of a Creative Commons Attribution License (CC-BY).Download (2.54MB)
In this work, the potential of carrying out inverse problems with linear and non-linear behaviour is investigated using deep learning methods. In inverse problems, the boundary conditions are determined using sparse measurement of a variable such as velocity or temperature. Although this is mathemat...
|Published in:||Computer Methods in Applied Mechanics and Engineering|
Check full text
No Tags, Be the first to tag this record!
In this work, the potential of carrying out inverse problems with linear and non-linear behaviour is investigated using deep learning methods. In inverse problems, the boundary conditions are determined using sparse measurement of a variable such as velocity or temperature. Although this is mathematically tractable for simple problems, it can be extremely challenging for complex problems. To overcome the non-linear and complex effects, a brute force approach was used on a trial and error basis to find an approximate solution. With the advent of machine learning algorithms it may now be possible to model inverse problems faster and more accurately. In order to demonstrate that machine learning can be used in solving inverse problems, we propose a fusion between computational mechanics and machine learning. The forward problems are solved first to create a database. This database is then used to train the machine learning algorithms. The trained algorithm is then used to determine the boundary conditions of a problem from assumed measurements. The proposed method is tested for the linear/non-linear heat conduction, convection–conduction, and natural convection problems in which the boundary conditions are determined by providing three, four, and five temperature measurements. This study demonstrates that the proposed fusion of computational mechanics and machine learning is an effective way of tackling complex inverse problems.
Inverse modelling, Computational mechanics, Machine learning, Heat conduction, Heat convection–conduction, Natural convection
College of Engineering