A recommendation engine for suggesting unexpected thermoelectric chemistries

I. INTRODUCTION

• Experimental efforts often favor incremental improvements to known chemistries over high-risk searches for new materials.Notable material classes have also been discovered fortuitously.Source: §I · p. 1
• Most investigated thermoelectrics cluster compositionally around chalcogenides and p-block elements such as Sn and Sb.The paper describes this homogeneity using materials in the Gaultois et al. database.Source: §I · p. 1
• Gd12Co5Bi and Er12Co5Bi are almost pure intermetallics, distinct from known thermoelectrics and proposed as a new class for further study.The engine’s objective is to help experimental researchers identify such candidates.Source: §I · p. 1

A materials recommendation engine

• The model assigns confidence that four room-temperature properties fall within specified ranges: Seebeck coefficient, resistivity, thermal conductivity, and band gap.The thresholds are |S| > 100 µV K−1, ρ < 10−2 Ωcm, κ < 10 W m−1 K−1, and Eg > 0 eV.Source: A materials recommendation engine · p. 1
• The engine identified RE12Co5Bi (RE = Gd, Er), which the paper characterizes as chemically and structurally distinct from known thermoelectrics.The compounds lie outside the existing composition space of common systems.Source: Figure 1 · p. 2
• A material meeting all four criteria is treated as potentially promising and worthy of further study, not as a quantitative prediction or a definitive record-setter.The authors describe quantitative property prediction and definitive identification of record-setting compounds as open challenges.Source: A materials recommendation engine · p. 2

Modelling and informatics

• Materials informatics applies empirical machine-learning methods to predict materials behavior, using training data, descriptors, and a learning algorithm.Descriptors vectorize materials through numerical or categorical characteristics, such as average atomic number or crystal structure.Source: Modelling and informatics · pp. 2–3
• The training set combines experimental thermoelectric characterization data with first-principles-derived electronic-structure data.The data include measurements such as Seebeck coefficients, thermal conductivities, electrical conductivities, and band gaps across materials and conditions.Source: Modelling and informatics · p. 3
• The engine blends in-house and sourced descriptors, encoding periodic-table knowledge such as similarities among lanthanides and Li and Na.The authors state that this pre-bias supports predictive models with datasets of thousands of examples.Source: Modelling and informatics · p. 3
• A random forest combines many decision trees trained on slightly different subsets of data to model complex nonlinear behavior.The paper gives electronegativity differences in NaCl as an example of a decision-tree rule relevant to band-gap prediction.Source: Modelling and informatics · p. 3

Model validation

• Leave-one-out cross-validation trains on n − 1 measurements and predicts the withheld value in turn, producing an error distribution used to estimate performance on unfamiliar materials.For band-gap data, the authors performed cross-validation on a subset of the large training set.Source: Model validation · p. 3
• The error distributions have modes close to zero, and the engine’s errors generally skew toward false negatives for resistivity, band gap, and thermal conductivity, but false positives for Seebeck.The authors attribute the directional skew to unequal proportions of positive and negative examples in the training data.Source: Model validation · p. 3
• Seebeck is the hardest property to assess, with the largest error-distribution standard deviation, which the authors link to differing underlying mechanisms.They added band gap as a secondary metric after earlier models produced too many good-metal candidates with poor Seebeck coefficients.Source: Model validation · p. 3

Experimental details

• RE12Co5Bi samples were arc-melted twice with 5–7% excess bismuth, sealed and annealed at 1070 K for one week, then quenched.Total mass loss after melting was less than 1%.Source: Experimental details · p. 3
• Figure 2 plots leave-one-out cross-validation error histograms for Seebeck coefficient, resistivity, thermal conductivity, and band gap; errors near zero indicate confident classifications close to the target ranges.Errors near +1 indicate false negatives, while errors near −1 indicate false positives.Source: Figure 2 · p. 4
• About 70 samples of each compound were combined into approximately 5 g ingots, with final pellet densities matching single-crystal values.The reported densities were 8.6 g/cm3 for Gd12Co5Bi and 9.9 g/cm3 for Er12Co5Bi.Source: Experimental details · p. 4
• Gd12Co5Bi was compositionally homogeneous and measured at Gd69(2)Co26(2)Bi5(2); Er12Co5Bi could not be quantitatively analyzed by EDX because of overlapping spectral lines.Powder X-ray diffraction and Rietveld refinement were used to confirm structure and phase purity.Source: Experimental details · p. 4
• Electrical resistivity and Seebeck coefficient were measured from 300 K to 800 K over three heating and cooling cycles.Measurements used a ULVAC ZEM-3 under helium under-pressure and lasted 18 hours to assess stability and reproducibility.Source: Experimental details · p. 4

III. DISCUSSION

• The web app lets researchers explore around 25,000 known compounds and evaluate their own candidates in real time.Users can sort by any of four modeled properties or a composite score and apply property cutoffs to narrow the list.Source: §III · pp. 4–5
• TaVO5 and TaPO5 are proposed as promising analogues because their heavy atoms, connected ReO3-type networks, and structural shear may favor thermoelectric performance.TaVO5 also has a structural transition at 600°C that may soften phonon modes and lower thermal conductivity; Tl9SbTe6 scores highly and later work demonstrated good thermoelectric performance [23].Source: §III · p. 5
• TaAlO4, SrCrO3, and TaSbO4 rank highly despite being expected insulators, partly because training references primarily report stoichiometric formulas rather than doping details [24, 25].Substitution or reduction may yield moderate electrical performance; SrCrO3 has been reported metallic when synthesized under pressure [26].Source: §III · p. 5
• The engine does not use crystal-structure features, which avoids requiring structures unavailable for uncharacterized materials but creates difficulty when structure changes alter properties at similar compositions.For example, predicted low thermal conductivity for both DyPO4 and LaPO4 misses their different structures and the inherently higher thermal conductivity of edge-shared xenotime DyPO4 [27, 28].Source: §III · p. 5

New materials and their properties

• The candidates were chosen from high-scoring materials partly because arc melting made them facile to synthesize and their chemistry differed from known thermoelectrics.The RE12Co5Bi family had been sparsely studied, with its crystal structure and initial low-temperature electrical and magnetic properties reported previously.Source: New materials and their properties · p. 5
• RE12Co5Bi has Ho12Bi icosahedra analogous to the LaP12 units in filled skutterudites, although the rare-earth and group 15 element roles are reversed.The authors hypothesize that this structural similarity could partly account for the compounds’ thermoelectric behavior; crystal structure was not an engine input feature.Source: New materials and their properties · p. 5, Figure 3 · p. 6
• Gd12Co5Bi and Er12Co5Bi are an unoptimized new thermoelectric class outside the usual search space, with resistivity comparable to chalcogenides but Seebeck coefficients too low for competitiveness.Their thermal conductivity ranges from 4 W m−1 K−1 to 8 W m−1 K−1 between 300 K and 800 K, comparable to half-Heuslers.Source: New materials and their properties · pp. 5–6, Figure 4 · p. 7
• The recommendation engine’s predictions agreed with the measured property thresholds, including low confidence for a large Seebeck coefficient.The electrical performance figure of merit κzT is around 0.03 W m−1 K−1 at 400 K, higher than nearly 30% of materials in the Gaultois thermoelectrics database.Source: New materials and their properties · p. 6, Figure 4 · p. 7
• The compounds’ thermal conductivity rises with temperature, despite expected decreases in both electronic and phonon contributions; the origin of the positive thermal-diffusivity dependence is unknown.Further study of this rare, though not unprecedented, behavior could lead to new thermoelectric optimization strategies.Source: New materials and their properties · p. 7

• Machine learning suggested an experimentally viable new compound from true chemical white space where no prior characterization hinted at promising chemistries.Source: §IV · pp. 7–8
• The approach uses a data-driven computational tool to augment experimental capabilities and intuition as a semi-rational route to materials families that may have desirable properties.Source: §IV · p. 8
• The authors suggest this paradigm could eventually replace trial-and-error and fortuity in searches for new materials across application areas.Source: §IV · p. 8