Data-driven approach to predicting heterogeneous nucleation in phase change materials

Abstract

There is a current requirement for technologies that store heat for both domestic and industrial applications. Phase-change materials (PCMs) are an important class of substances with strong potential for heat storage. For practical use, storage systems must withstand repeated melt/freeze cycles while maintaining a stable melting-crystallisation point and consistent heat output. Salt hydrates are attractive candidates on account of their high energy densities, but there are issues associated with its strong tendency to subcool well below its normal freezing point. While the nucleation problem can be readily solved by the addition of seed crystals of another material, there are a lot of problems that can be encountered that result in nucleator deactivation. Therefore, the problem of identifying suitable heterogeneous nucleating crystallites (NUCs) for PCMs under variable temperature conditions remains a challenging task. In this regard, in silico screening methods offers a practical solution to both problems. Through a data driven approach, a workflow is generated by learning from existing experimental reports of working PCM/NUC pairs, in the light of searching for other NUC candidates that may offer improved properties over the additives that are currently used. The focus of this research is therefore to demonstrate the feasibility of a data-driven approach to establish a high-throughput NUC prediction model that could be applied to any given liquid/solid PCM.

In Chapter 2, a workflow generation process is described. The workflow is based on a data-driven approach, and a high-throughput workflow is created based on geometric matching under five related features that returns a binary decision of working/non-working NUC for a given PCM.

In Chapter 3, the trained model is applied with a most extensively studied PCM, ice-Ih. The model is firstly utilised to evaluate the degree of nucleation effectiveness then compared with already existing experimental reports. Bulk water immersion experiments on a set of ten known nucleators sets a delineating temperature to distinguish between good and poor nucleation behaviour. The algorithm is then used to screen 3,500 simple metal oxides and halides taken from the Inorganic Chemistry

Structural Database (ICSD), and show that just 7% of the former and 3% of the latter were predicted to nucleate ice on the basis of geometric slab matching. Subsequent experimental testing of 22 compounds suggests a 64% correct prediction rate, and identifies four new ice nucleators. Inspired by the ice-nucleating efficiency of copper oxides, the copper tubing with local tap water is also tested, and subcooling suppression is observed, most likely due to copper oxide build-up.

In Chapter 4, the model is further trained and tested with working/non-working nucleators from readily existing reliable experimental reports and then the trained model is applied in a high-throughput application for sodium acetate trihydrate (SAT), where over 14,000 candidate NUC structures are screened, from which a list of 521 compounds is identified as potential NUCs for SAT. The result reinforces the success of the current industry-standard NUC for SAT, disodium hydrogen-phosphate hydrates (DSP), which is shown to geometrically match slabs of SAT regardless of the level of hydration present. Other PCMs are sought after, i.e. Mg(NO3)2∙6H2O, MgCl2∙6H2O, CaCl2∙6H2O, and LiNO3∙3H2O. The distribution of prediction from working to non-working NUCs for the four PCMs demonstrates mostly the same trend as the confidence range. This meant this model could be readily used for nucleator mining for other PCM materials.

In Chapter 5, a supervised machine learning workflow is set up with the goal of predicting effective nucleators for any PCM material based on geometric compatibility between their crystallographic slabs. The algorithm is established by learning from highly granular geometric data generated from ice nucleation in Chapter 3 and salt hydrates in Chapter 4, and this approach avoids manually tuning thresholds and instead lets the model discover which geometric criteria (and value ranges) are statistically associated with successful nucleation. The results show prominent prediction power, i.e. success rate on both ice and salt hydrates, and further data analysis showed equal contributions as well as independence of the five features, proving the comprehensiveness of the algorithm.

The impact of this research as well as future works are discussed in Chapter 6.