Jie Gong1,Sharon Chu1,Rohan Mehta1,Alan McGaughey1
Carnegie Mellon University1
Jie Gong1,Sharon Chu1,Rohan Mehta1,Alan McGaughey1
Carnegie Mellon University1
<b>Electrocaloric (EC) cooling offers great potential to build efficient solid-state cooling devices that are quiet, low weight, and compact. The key to building practical EC devices is a large EC temperature change. However, identifying an effective EC material is a non-trivial task. The exploration of new EC materials relies on the instincts of experts and extensive experimental synthesis of ceramics, polymers, and/or composite materials. To aid the search for effective EC materials, we apply a physics-informed data-driven approach to build an eXtreme Gradient Boosting (XGBoost) model to predict the EC temperature change for ceramics based on the material composition (encoded with the Magpie package by elemental chemical properties), dielectric constant, Curie temperature, and characterization conditions. The dataset consists of 97 EC ceramics gathered from published experimental literature. The model achieves a coefficient of determination score (i.e., R2 score) of 0.78 for the test data. The feature analysis shows that the model captures the known physics for effective EC materials (i.e. large applied electric fields and characterization temperatures around Curie temperature). Although the Magpie features synergistically help the model distinguish between different materials, features related to electronegativity and ionic charge of individual elements are considered more important than others by the XGBoost model. To further improve the performance, structural information such as sample size, morphology, and grain size would be needed. We then apply the model to search for effective EC materials from 66 ferroelectrics whose EC performance has not been characterized. Effective lead-free candidate materials (EC temperature change above 2 K predicted around room temperature and at an electric field of 100 kV/cm) are suggested for future experimental verification. Our model demonstrates the effectiveness of physics-informed machine learning, where domain-specific knowledge is leveraged with advanced algorithms.</b>