Key points are not available for this paper at this time.
ABSTRACT The Seebeck coefficient is an intrinsic thermoelectric material property, and is a key parameter in microelectromechanical systems (MEMS) sensors and energy harvesting based on the thermoelectric principle. However, extracting this coefficient accurately from MEMS test structures demands a careful co‐design to balance measurement fidelity against power consumption. Till now, the trade‐off between accuracy and power consumption is not very well addressed by traditional approaches based on computationally expensive finite element analysis (FEA) or trials and errors methods. To address this issue, we present a physics‐guided deep learning (PGDL) surrogate model that predicts key device level parameters; output voltage, heater resistance and temperature of the hot and cold junctions through model training on 5000 high‐fidelity data‐set generated on COMSOL Multiphysics simulations. Rigorous benchmarking against conventional deep learning, random forest, k‐nearest neighbors, and gradient boosting, confirms that PGDL consistently delivers superior predictive accuracy and generalization. In PGDL, domain‐specific physical knowledge for instance, expected monotonic trends between geometry and thermal response are integrated in a weakly constraint form during training, improving generalization and avoiding non‐physical predictions without imposing a form of a partial differential equation. We subsequently combine this surrogate with the NSGA‐II multi‐objective optimizer and use it to minimize (i) the relative error in recovering a known reference Seebeck coefficient, and (ii) the heater power consumption subjected to constant current operation concurrently elucidating the trade‐offs between conflicting objectives. The achieved Pareto optimal design consequently realizes a power dissipation drop by 48% (1.63 to 0.85 mW ), and minimize the Seebeck prediction error up to 2.1%–2.96%, while allowing an R 2 values between 0.948 and 0.970 across all model predicted outputs. This research demonstrates a scalable, data efficient, and physically consistent computational framework for the automated co‐design and in situ characterization of MEMS thermoelectric devices.
Ullah et al. (Wed,) studied this question.