Machine Learning Strategies for Screening Mn-T-X (T=Co, Fe, Ni; X=Si, Ge, Sn) Multicaloric Materials

Abstract

Solid-state cooling devices based on the “caloric” class of functional materials are considered a promising alternative to conventional vapor compression technologies as they allow complete elimination of high-global warming potential (GWP) refrigerants and have high energy efficiency that is theoretically equivalent to 60% of Carnot efficiency. [1] By definition, multicaloric materials exhibit reversible thermal changes that can be driven concurrently or in sequence by more than one type of external energy sources (magnetic field, electric field, strain/pressure). The use of multiple driving forces can bring about larger thermal changes with smaller field magnitudes over broader operating temperature ranges, while reducing energy losses due to hysteresis in one control parameter by shifting it to another.Multicaloric materials under current investigation are fraught with issues related to the inclusion of strategically-limited and toxic elements (e.g., Gd5(GeSi)4, FeRh, MnAs; bold=problematic element) or require complex synthesis and processing to realize an acceptable, or even marginal, functional response. Against this backdrop, the MnTX family of compounds [T = transition metal elements (Fe, Ni, or Co); X= main group p-block element (Si, Ge, or Sn] are poised to overcome these limitations since they are made of earth-abundant, non-toxic elements, are scalable for powder production using low-cost, conventional solid-state processing techniques. Select MnTX alloys demonstrate tunable, room-temperature, first-order magnetic phase transition (i.e., magnetostructural transition from a low temperature TiNiSi-type ferromagnetic orthorhombic phase to a high temperature Ni2In-type paramagnetic hexagonal phase) with promising magnetic fields and pressure-driven caloric responses. Though one can tune the magnetic and caloric properties through careful choice of chemical elements that occupy T- and X-sites or their relative concentrations, large chemical diversity presents formidable combinatorial challenges for experimental and theoretical studies.To accelerate the discovery of new MnTX materials with low hysteresis magneto-structural phase transitions near room temperature, we apply a machine learning (ML) approach utilizing data from published literature to construct data-driven ML models as illustrated in figure one. To this end, we start by performing a literature review and constructing a database representative of the MnTX materials that undergo the Pnma to P63/mmc magneto-structural transition. [2]-[21] We found 87 unique observations that have been experimentally synthesized and whose Pnma to P63/mmc transition temperature on heating (TH) and P63/mmc to Pnma transition temperature on cooling (TC) have been characterized. The difference between |TH – TC| is indicative of the hysteresis associated with the phase transition. From these 87 observations, we created 2 different methods of representing the data for machine learning. The first was created using element specific features (e.g., valence electron concentration, electronegativity, and size), where each composition is described as a weighted average of these individual elemental properties. In the second method, we performed density functional theory (DFT) to calculate certain properties (e.g., magnetic moment, bond lengths, and unit cell parameters) of end-point compositions for each compound using the open source electronic structure code quantum espresso.[22]–[24] Each alloy or solid solution was then described as a weighted average of the properties of each end point composition. Models were trained using bootstrapped LASSO, support vector regression (SVR), gaussian process regression and random forest methods implanted in the statistical computing language, R.[25] Of these algorithms SVR was the most accurate on the validation set for both the elemental feature space and the DFT feature space. The perfomance of this model is shown in figure two. When compared to each other the model trained on the DFT generated features out-performed the model trained on the features generated from elemental properties. Promising candidates are recommended for experimental validation and feedback for iterative model improvement. **