Machine learning for thermoelectric material discovery
Researchers in China have developed a machine learning (ML) algorithm to streamline the discovery and development of high-performance ionic thermoelectric (iTE) materials.
The ML framework provides a rapid and precise method for predicting Seebeck coefficients for thermoelectric materials for sensors and heat recovery systems.
Ionic thermoelectric materials are known for their high Seebeck coefficients, often surpassing 10 mV/K, two orders of magnitude higher than their electronic counterparts. However development of thermoelectric engines can be labour-intensive with trial-and-error experimentation.
The iTE materials experience uneven anion and cation diffusion under temperature gradients, generating an electrical potential difference to power external devices.
A ‘simplified molecular-input line-entry system’ (SMILES) developed by the team at Tsinghua University in China Is used for encoding molecular structures, enabling the machine learning model to handle the diversity of iTE material types effectively.
A ML regression model is trained with the features to evaluate the Seebeck coefficient accurately for direct screening of i-TE materials.
The ML model is trained on a carefully curated dataset of 51 i-TE material samples, achieved an impressive coefficient of determination (R²) of 0.98 in test cases. By analyzing molecular features derived from SMILES, the model predicted a range of promising materials, among which a waterborne polyurethane-potassium iodide (WPU/KI) ionogel stood out. Experimental validation confirmed the ionogel’s Seebeck coefficient at 41.39 mV/K.
Through interpretable analysis, the team identified critical molecular descriptors influencing the Seebeck coefficient. Two features stood out: the number of rotatable bonds and the octanol-water partition coefficient of ion donors. Molecular dynamics simulations further confirmed their impact, revealing that these properties govern ion diffusion and interaction with the polymer matrix—key mechanisms behind high thermopower. For instance, a higher number of rotatable bonds correlated with reduced ion diffusion efficiency, while a low partition coefficient indicated stronger Coulomb interactions, enhancing thermoelectric performance.
This study also highlights the complementarity of the ML models. While gradient boosting decision trees (GBDT) provided high-precision predictions within known ranges, symbolic regression models like genetic programming symbolic regression (GPSR) demonstrated utility in extrapolating beyond the dataset, suggesting novel material combinations.
The researchers validated their predictions by fabricating and testing the top-ranked materials. The WPU/KI ionogel exhibited stable performance under various conditions, with a temperature-induced voltage increase of 250 mV for a gradient of 5.5 K. Molecular dynamics simulations illuminated the physical interactions driving this performance, such as the strong electrostatic attractions between potassium ions and the polyurethane matrix.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
