Machine Learning for Data Assimilation/Inverse Modelling in Thermal Problems
This study aims to systematically examine various methods leveraging machine learning (ML) to assimilate data for investigating thermal systems. These measured or observed data may include temperature or thermal material properties and could be synthetically (computationally) generated or experimentally obtained. The goal of these ML-augmented methods is to derive the unknown material properties and/or reconstruct the full temperature field by integrating such measured data into physics-based computational models, such as FE models. The present work continues the previously conducted review of ML in heat transfer with a strong focus on inverse modeling techniques. It also attempts to closely incorporate ML into the FE workflow. Data assimilation and inverse modeling are closely linked tasks. While inverse modeling typically focuses on recovering unknown parameters or inputs from given observational data, data assimilation incorporates observations into dynamic models in a sequential manner, often with the goal of improving forecasting performance. In this review, we use the terms interchangeably for simplicity, though they arise from distinct methodological traditions.