Artificial Neural Networks and Deep Learning in the Design and Applications of Phase Change Materials: A Review

Artificial neural networks (ANNs) and deep learning (DL) are increasingly used to accelerate the design of phase change materials (PCMs), yet existing surveys catalogue these applications without assessing whether the architectures used suit the structure of their input data. As a result, the high accuracies reported in this literature say little about how models will perform beyond their training data. This review evaluates 40 studies across four architectural families (multilayer perceptrons, convolutional neural networks, recurrent networks, and hybrid models), covering the prediction of material properties, transient heat transfer, and system-level performance in battery thermal management, buildings, and solar energy. The analysis yields three main findings. First, accuracy is mostly measured within the training distribution and therefore reflects interpolation rather than predictive capability. Second, architectural choices are rarely justified by input structure: convolution is often applied to vectors of independent scalars, and the need for recurrence is seldom tested against a feedforward baseline. Third, where such comparisons exist, simpler models are frequently competitive. In practice, researchers should match each architecture’s inductive bias to a structure demonstrably present in its input, report a simple baseline, and validate models beyond the training range.