Thermoelectric screening moves from single points to whole curves: composition plus temperature yields zT(T) with held-out R² of 0.88
Synopsis
Thermoelectric screening needs the full zT-versus-temperature curve, because peak height and position set the usable operating window, yet measuring S, σ and κ point by point is costly. A multi-task model with zT as the principal task and S, σ, κ and PF as auxiliary tasks takes only composition descriptors and temperature, predicts all five properties and reconstructs zT from them. In ten-fold material-grouped held-out evaluation zT reaches R² = 0.8765 with peak-temperature MAE of 14.07 K, and 24 external points from four literature materials give R² = 0.9755.
Interpretation
Given a chemical formula and a temperature, the model returns the Seebeck coefficient, electrical conductivity, thermal conductivity, power factor and figure of merit zT at that temperature, and from these traces the whole zT-versus-temperature curve. Thermoelectric transport curves previously required measuring S, σ and κ over dense temperature intervals, often with different instruments or sample conditions; first-principles and Boltzmann-transport calculations offer microscopic insight but their cost and structural requirements limit broad screening across realistic alloying, defects and carrier concentrations. The curated dataset holds 4222 observations from 755 formula-defined materials and 270 elemental systems over 275–1275 K; ten-fold material-grouped out-of-fold prediction gives zT R² = 0.8765, MAE = 0.0764 and RMSE = 0.1166, with per-fold R² between 0.8499 and 0.9291.
The model treats zT as the principal output and the other four properties as auxiliary supervision, using the thermoelectric relations to keep the outputs mutually consistent during training. Multi-task learning lets related targets share information, but the five targets differ in scale, noise, coverage and relevance to screening, so equal weighting spreads capacity onto the auxiliary quantities. Composition enters as 132 Magpie descriptors filtered by variance and correlation and reduced by principal component analysis to 96 dimensions, while temperature is expanded with four Fourier frequencies; training runs in two stages, the second raising the zT loss weight from 6.00 to 8.00 and lowering auxiliary weights, with high-zT reweighting and a physical-consistency penalty ramped in over the first 30 epochs.
Curve-level evaluation isolates peak height and the temperature at which the peak occurs, rather than reporting pointwise error alone. Random pointwise splitting places different temperatures of the same material in both training and test sets, so it measures interpolation along partially observed curves rather than prediction for unseen materials. With materials grouped, peak-zT MAE is 0.1141 and peak-temperature MAE is 14.07 K; predictions for eight representative held-out materials follow the measured direction of change, with trend MAE of 0.027–0.136 per 100 K.
Perspective
The model targets screening settings where only a chemical formula is known and no detailed crystal structure is available, producing temperature-dependent property curves over 275–1275 K for candidate ranking and operating-window selection. Curve-level analysis includes only materials with at least four distinct experimental temperatures, while pointwise evaluation uses all valid material-temperature observations. External validation covers four literature-derived materials, BaSnS3, Na2InCuF6, LaOBiS2 and MgI2Bi2S3, totaling 24 temperature-zT points.
External validation involves only four materials and 24 points, with material-level R² between 0.9347 and 0.9878, so the sample is limited. Across the ten material-grouped folds zT R² ranges from 0.8499 to 0.9291, so transfer between material groups is uneven. The composition representation encodes formula information only; structure, carrier concentration, microstructure, synthesis history and measurement conditions stay outside the model and serve as additional context for material-specific refinement.