Closed-Loop Error-Correction Learning Accelerates Experimental Discovery of Thermoelectric Materials.

Area of Science:

• Materials Science
• Solid State Physics
• Computational Materials Science

Background:

• Discovering novel thermoelectric materials is complex due to vast material combinations, doping possibilities, and synthesis methods.
• Traditional high-throughput screening methods, even with advanced machine learning (ML), face challenges in adapting to synthesis and characterization variations.

Purpose of the Study:

• To develop and apply an error-correction learning (ECL) strategy for efficient thermoelectric material discovery.
• To prioritize material synthesis at temperatures below 300 °C.
• To identify novel thermoelectric material families and optimize their performance.

Main Methods:

• Incorporation of historical data and iterative refinement using experimental feedback via ECL.
• Adaptation of ML models to account for synthesis and characterization variability.
• Closed-loop experimentation strategy focused on low-temperature (<300 °C) synthesis.

Main Results:

• Identification of a new thermoelectric material family: PbSe doped with SnSb.
• The optimized material, 2 wt% SnSb doped PbSe, demonstrated a power factor more than double that of pure PbSe.
• The closed-loop strategy reduced the number of experiments by up to 3× compared to standard ML-driven high-throughput searches.

Conclusions:

• Error-correction learning (ECL) combined with closed-loop experimentation significantly accelerates thermoelectric material discovery.
• The efficiency gains are substantial, particularly when ML model accuracy reaches a certain threshold, after which experimental pathway optimization becomes critical.