Augmenting machine learning to thermoelectric measurements : using Bayesian inference to infer electronic transport parameters from device level power-load data

Abstract

Thermoelectric materials efficiency is characterized by Figure of Merit. Figure of Merit depends on several thermoelectric descriptors which is thermal conductivity and Power Factor, which is equal to the product of electrical conductivity and square of Seebeck coefficient. Hence, scientists have been looking for a rapid characterization of thermoelectric for highput materials synthesis. In this project, we use Bayesian inference algorithm on power-load experiment dat to infer the transport properties of thermoelectric materials. We focused on conventional Bi-Sb-Te thermoelectric materials,Bi0.4Sb1.6Te3, Bi0.5Sb1.5Te3, and Bi0.6Sb1.4Te3. These samples are prepared using ball miling followed by Spark Plasma Sintering (SPS) process. We obtained the power out of a single leg as function of load resistance at base temperature of 300 K in ultra-high vacuum in a closed-cycle He cryostat. The experiment is repeated under different temperature gradients accros the samples as discriminative testing condition for the machine learning algorithm to infer the electrical transport properties. Seebeck coefficient (S), electrical conductivity (), contact resistance (Rc), and error in temperature difference are inferred and presented in probability distribution. Along with these parameters, Bayesian inference helps to estimate the intrinsic properties of thermoelectric materials, such as energy depedent scattering, doping and weighted average mobility, which are very tedious to measure experimentally.