Zone Method meets Physics-Informed Neural Networks: Data and Regularization for High-Temperature Processes
Abstract: In this paper, we visit the classical Hottel's zone method from a Machine Learning (ML) perspective. The zone method provides an elegant, iterative way to model high temperature processes, such as in industrial furnaces, which are energy-intensive components in the Foundation Industries (FIs). Real-time and accurate temperature modeling in reheating furnaces can help the FIs achieve their Net-Zero goals for the greater good of environmental sustainability. As computational surrogate models are expensive, and slower, we propose to leverage Deep Learning (DL) owing to their effectiveness and inference efficiency. However, obtaining quality large-scale data for DL/ML training in systems with high temperature processes such as furnaces is challenging. To address this, we propose an algorithm to use the zone method for generating data for training ML/DL models via regression. But, DL inherently finds it challenging to generalize to Out-of-Distribution (OOD) samples. Thus, we additionally propose to employ zone method inspired novel Energy-Balance (EB) based regularizers to explicitly convert a neural network to a Physics-Informed Neural Network (PINN) for adhering to the underlying physical phenomenon during network training, to be able to better handle OOD data. We support our claims with extensive empirical experiments, thus showcasing the promise of our PINN. Our PINN framework has been demonstrated for both Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) based Recurrent Neural Network (RNN) architectures.
Submission Length: Long submission (more than 12 pages of main content)
Assigned Action Editor: ~Ole_Winther1
Submission Number: 1869
Loading