A low-temperature, water-free fabrication route to Mg-based micro thermoelectric coolers for thermal management

Abstract

Sustainable thermal management of high-heat-flux electronics beyond Moore’s Law requires micro thermoelectric coolers (μ-TECs) that are both high-performance and environmentally sustainable. However, state-of-the-art μ-TECs rely on Bi₂Te₃-based materials, whose tellurium scarcity, toxicity concerns and limited mechanical robustness hinder widespread adoption. Mg-based thermoelectric materials offer a promising alternative but remain challenging to integrate into microscale devices. Here we show that a low-temperature, water-free fabrication strategy enables tellurium-minimized μ-TECs based on n-type Mg₃(Bi,Sb)₂ and p-type MgAgSb. Using magnetron sputtering as a cold bonding approach, we fabricate compact 12-pair microdevices (2.95 × 4.35 × 1.4 mm³), with thermoelectric legs approximately 3% the size of previously reported Mg-based devices. The resulting μ-TECs achieve a power density of 4.34 W cm⁻² and a packing density of 93.5 pairs cm⁻². Our work establishes a scalable route for sustainable thermoelectric cooling and a viable alternative to conventional Bi₂Te₃ technologies.

This study reports Mg-based micro thermoelectric coolers fabricated via a low-temperature process, achieving high cooling performance and enabling a scalable, sustainable alternative to conventional Bi2Te3-based devices for thermal management.

Introduction

The relentless miniaturization of high-heat-flux electronic systems in the post-Moore’s Law era has placed stringent demands on sustainable thermal management technologies, with microthermoelectric coolers (μ-TECs) emerging as a key solution for localized, energy-efficient cooling. Conventional Bi₂Te₃-based μ-TECs, however, face insurmountable barriers to widespread application: tellurium’s inherent toxicity and resource scarcity conflict with global sustainability goals, while the material’s mechanical fragility renders it incompatible with precise microfabrication processes. In response, n-type Mg₃(Bi, Sb)₂ and p-type MgAgSb have emerged as promising tellurium-free alternatives, but their translation into microdevices has remained elusive due to the materials’ extreme sensitivity to moisture and limited compatibility with high-temperature processing. To bridge this gap, we develop a low-temperature and water-free route that circumvents these intrinsic material limitations, enabling the fabrication of high-performance Mg-based μ-TECs and unlocking their potential for electronics.

Thermoelectric materials enable direct conversion between heat and electricity, providing scalable solid-state solutions for power generation and active cooling^1,2. Their microscale implementation—micro-thermoelectric coolers (μ-TECs)—delivers exceptional cooling power density and localized thermal control, addressing critical heat management challenges in optoelectronic and microelectronic systems^3,4. Current μ-TECs predominantly use Bi₂Te₃, a material plagued by mechanical fragility and environmental concerns due to its tellurium content—specifically because of Te’s toxicity and low terrestrial abundance, limiting further device miniaturization^3–5. While numerous high-zT materials (e.g., Mg₃(Bi,Sb)₂^5–7, MgAgSb^8–10, SnSe^11–13, PbSe^14,15, AgSbTe₂^16,17, Ag₂Se^18–20, and half-Heuslers²¹) have been developed, their integration into functional microscale devices remains a key bottleneck.

Developing μ-TECs from emerging materials faces two major challenges. First, the R&D cycle for device fabrication far exceeds that for material synthesis. Miniaturization—for instance, reducing Mg₃(Bi,Sb)₂ leg volumes from macro scales (>1.5 × 1.5 × 2  mm³) to μ-TEC dimensions (~0.3 × 0.3 × 0.4 mm³)—dramatically increases difficulties in positioning, cutting, and soldering^3,22. Second, interfaces for emerging thermoelectric materials are typically prepared via one-step hot-pressing, producing thick, non-uniform layers under manual control, thereby hindering independent optimization of thermoelectric and interface properties^9,10,22–25. This is particularly critical for materials like Mg₃(Bi,Sb)₂ and MgAgSb, under high-temperature sintering conditions, the volatilization of ingredients, phase transition, and material decomposition are accelerated. Moreover, the high-temperature treatment used to improve grain size and carrier mobility also accelerates interfacial diffusion^26–29. Addressing these issues requires both theoretical interface design and advanced fabrication strategies.

We report the development of high-performance, tellurium-minimized μ-TECs based on n-type Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 and p-type MgAgSb for near-room-temperature applications. Employing room-temperature cold interface bonding via magnetron sputtering, a low-temperature and water-free route, we fabricated Cu/Mo and Ag/Mo multilayer interfaces on Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 and MgAgSb, achieving robust adhesion and uniform morphology; notably, the Cu/Mo/Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 junction exhibited an ultralow interfacial resistivity of 5.9 μΩ cm². Leveraging this strategy, we demonstrate a tellurium-minimized 12-pair Mg-based μ-TEC that is only 3% the size of previously reported normal-sized devices (leg size shrinking from ~1.5 × 1.5 × 2 mm³ to 0.48 × 0.48 × 0.56  mm³) and a packing density of 93.5 pairs cm⁻², delivering a maximum cooling power density of 4.34 W cm⁻²—a fourfold enhancement over prior Bi₂Te₃-free Mg-based devices. These results highlight the potential of environmentally friendly, high-performance miniaturized thermoelectric modules to advance sustainable thermal management in electronics and provide a route towards the miniaturized integration of emerging thermoelectric materials.

Results

Interfacial engineering technology for Mg-based materials

We first conducted a comprehensive evaluation of interface fabrication techniques—including hot pressing (HP), electroplating (EP), chemical vapor deposition (CVD), and sputtering—highlighting differences in operating temperature and the ability to decouple the synthesis of thermoelectric materials and thermoelectric interface materials (TEiMs), as shown in Fig. 1a. For microscale Mg₃(Bi,Sb)₂ and MgAgSb devices, fabricating a sandwich structure consisting of interface powders or foils and thermoelectric legs via HP poses a major limitation: poor control over interface thickness and uniformity. Layers often reach several hundred micrometers—comparable to the leg length—and post-sintering treatments like polishing or cleaning exacerbate non-uniformity, hindering precise device fabrication^{22–25,27,28}. Notably, achieving simultaneous elemental diffusion suppression, strong interfacial bonding, and high thermoelectric performance remains challenging, as high-temperature processing during HP frequently induces variability and instability during thermoelectric leg fabrication.

Fig. 1. Comparison of interface preparation techniques and the performance of Mg₃(Bi,Sb)₂/MgAgSb-based μ-TEC.

a Comparison of operating temperatures for various interface fabrication techniques, including hot pressing (HP), electroplating (EP), chemical vapor deposition (CVD), and sputtering^{5,9,24,32–38}. b Cross-sectional SEM image of randomly selected Cu/Mo/Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 and Ag/Mo/MgAgSb junctions, prepared by ion beam cutting (IBD). c Comparison of the contact resistivities (ρ_c) and layer thicknesses of commonly used interface materials for Mg₃(Bi,Sb)₂^{22,25,27,30,39–42}, with thickness here taken as double, as each leg is coated with interface layers on both sides. d Photograph of an Mg₃(Bi,Sb)₂/MgAgSb-based μ-TEC, with module dimensions of 2.95 × 4.35 × 1.4 mm³. e Comparison of maximum cooling power density (q_cmax) and packing density for recently reported thermoelectric coolers based on emerging materials such as Mg₃(Bi,Sb)₂^22,44–47, MgAgSb^45,46, Ag₂Se^48,49, and PbSe^14,15.

High room-temperature performance in Mg₃(Bi,Sb)₂ requires sintering at ~973 K for grain growth and reduced grain boundary scattering, but high-temperature processing induces elemental diffusion that degrades thermoelectric performance^26,29. Fe, a commonly used TEiM for Mg₃(Bi,Sb)₂^5,27, was shown by Qu et al. to cause a 75% conductivity drop and n-type-to-p-type conduction reversal when an Fe/Mg₃(Bi,Sb)₂/Fe stack was hot-pressed at 1073 K²⁷. Similarly, Wang et al. reported reduced power generation efficiency in single-leg Ni/Mg₃(Bi,Sb)₂/Ni devices after annealing at 573 K³⁰. In MgAgSb, the α-to-β-phase transition near 573 K induces abrupt carrier transport changes⁸, making hot-pressed interface layers with high-melting-point metals (such as Ti, V, and Cr) difficult to fabricate³¹. Typically, the hot-press sintering temperature is maximized without exceeding the melting or phase-transition point of the thermoelectric substrate—for example, ~700 K for Bi₂Te₃/Ni³², ~800 K for PbTe/Co_0.8Fe_0.2³³, ~900 K for Mg₃(Bi,Sb)₂/Fe, Co or Cr^5,34, and below 573 K for MgAgSb/Ag, Co, or MgCuSb^9,24. Interface-layer sintering temperature’s dependence on the substrate underscores the need to decouple thermoelectric material fabrication from TEiM fabrication. Unlike hot pressing, thin-film deposition methods enable decoupled optimization of thermoelectric materials and interface layers—opening a viable path to fabricate interface layers for microscale devices. However, variations in processing temperature and chemical reactivity restrict the suitability of some techniques. Apart from its high processing temperature, CVD is further constrained by complex gas handling and limited precursor availability, which hinder process control and may trigger side reactions³⁵. EP operates at low temperatures (<373 K) but risks corrosion from acidic/basic electrolytes³⁶. Compared with conventional HP- and EP-based interface fabrication strategies, sputtering offers a more suitable route for Mg-based micro-TECs because it enables low-temperature, solvent-free, and thickness-controllable interface construction^37,38. Although sputtered interfaces have been widely explored in Bi₂Te₃-based devices⁴, their application to Mg-based thermoelectric materials remains scarce, and direct transfer of these established interface systems is limited by the distinct physicochemical properties and interfacial stability requirements of Mg-based materials. Accordingly, we adopted magnetron sputtering to develop material-specific interfacial layers for Mg-based thermoelectric materials, addressing a rarely explored interface-fabrication route for Mg-based micro-TEC integration.

To implement this strategy, we deposited bilayer thermoelectric interfaces via room-temperature cold magnetron sputtering: Mo/Cu on n-type Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 and Mo/Ag on p-type MgAgSb. Focused ion beam (FIB) cross-sections of randomly selected regions (Fig. 1b) reveal compact, uniform structures with distinct layering and strong adhesion to thermoelectric substrates; their thicknesses are ~0.5/1 μm (Mo/Cu) and 0.5/2 μm (Mo/Ag), respectively. For comparison, Fig. 1c summarizes the combined thicknesses of the interface layers (hot-pressed) at both hot and cold sides—along with their contact resistivities—for Mg₃(Bi,Sb)₂. Our sputtered Mo/Cu bilayer (fabricated in this work) exhibits a contact resistivity of 5.9 μΩ cm²—markedly lower than that of hot-pressed pure metals like Fe and Ni, and comparable to values reported for multicomponent alloys^{22,25,27,30,39–42}. Notably, its thickness is nearly two orders of magnitude smaller than hot-pressed counterparts (often >200 μm), whose thicknesses approach the leg heights of mini-bulk or macro-scale thermoelectric modules^{22,25,27,30,39–42}. By contrast, our sputtered layers are only a few micrometers thick—comparable to the leg lengths of MEMS-fabricated micro-film modules⁴³. Building on this optimized interface strategy, we fabricated 12-pair Mg₃(Bi,Sb)₂/MgAgSb-based μ-TECs with ultra-compact thermoelectric legs of 0.48 × 0.48 × 0.56 mm³ (Fig. 1d). Specifically, the overall module dimensions are 2.95 × 4.35 × 1.4 mm³, resulting in a high leg fill factor of 43.1% and a packing density of 93.5 pairs cm⁻². Critically, our μ-TEC achieves a maximum cooling power density ($q_{c,\max }$) of 4.34 W cm⁻²—roughly fourfold higher than recent Mg₃(Bi,Sb)₂- or MgAgSb-based devices reported in literature^44–47, and over three times higher than PbSe- and Ag₂Se-based modules^14,15,48,49—with an overall packing density 4–9 times greater than those in previous reports (Fig. 1e).

Cold interfacial bonding

To enable cold interfacial bonding, we identified Cu as a suitable sputtered interfacial material because its thermal expansion coefficient (~20 × 10⁻⁶ K⁻¹) closely matches Mg₃Sb₂ (~17 × 10⁻⁶ K⁻¹), and Cu/Mg₃(Bi,Sb)₂ interfacial bonding strengths up to 43 MPa have been reported^22,34. Similarly, Ag and MgAgSb share comparable expansion coefficients (20 vs. 21 × 10⁻⁶ K⁻¹), and Ag forms strong bonds with MgAgSb via hot pressing^9,10. Accordingly, we first attempted Cu/Mg₃Bi_1.5Sb_0.5 and Ag/MgAgSb interfacial structures, depositing layers via room-temperature magnetron sputtering. However, scanning electron microscopy (SEM) revealed weak bonding at the Cu/Mg₃Bi_1.5Sb_0.5 interface, with numerous voids (Supplementary Fig. 1), while the Ag/MgAgSb interface showed even poorer adhesion and could be manually delaminated (Supplementary Fig. 2). Given such poor interfacial bonding, we turned to the literature to identify a viable solution. Wu et al. studied element diffusion in Mg₃(Bi,Sb)₂ annealed at 500 °C for 15 min, showing that Cu and Ni diffuse relatively rapidly, whereas Mo, Cr, Nb, and W exhibit much lower diffusivity³⁴. However, their high-melting points hinder effective bonding with Mg₃(Bi,Sb)₂ at thermoelectric-material-compatible sintering temperatures. We therefore used Mo, which combines low diffusivity with a relatively low work function and higher sputtered-particle energy than Cu and Ag⁵⁰, as a transition interlayer in a bilayer interface strategy: introducing Mo as a transitional interlayer to form Cu/Mo/Mg₃Bi_1.5Sb_0.5 and Ag/Mo/MgAgSb interfaces for the thermoelectric legs. A Cu/Mo bilayer was sputtered onto an Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 pellet (ϕ 12.7 mm, thickness 2–3 mm; Supplementary Fig. 3a). SEM and energy-dispersive spectroscopy (EDS) analyses confirm high interfacial uniformity, density, and structural continuity (Supplementary Fig. 3b, c). For analysis, a polished cross-section of the Cu/Mo/Mg₃Bi_1.5Sb_0.5 interface was examined via SEM and EDS, revealing that the Mo interlayer significantly enhanced interfacial integrity—no voids were detected at either the Cu/Mo or Mo/Mg₃Bi_1.5Sb_0.5 interfaces (Supplementary Fig. 4). High-angle annular dark-field (HAADF) scanning transmission electron microscopy (STEM) images of the Cu/Mo/Mg₃Bi_1.5Sb_0.5 junctions, combined with elemental mapping, reveal a uniform distribution of elements across the Cu, Mo, and Mg₃Bi_1.5Sb_0.5 layers without any obvious segregation (Fig. 2a). The Mo and Cu layers are clearly distinguishable, with thicknesses of ~0.5 μm and ~1 μm, respectively. Elemental line scans reveal sharp compositional transitions at the interfaces without plateau regions (Fig. 2b), indicating the absence of intermediate reaction products. Consistently, the Mg-Bi-Mo ternary phase diagram shows no Mo-containing compounds within this system, apart from Mg₃Bi₂ and Mg₁₄₉Bi (Supplementary Fig. 5). The HAADF-STEM and bright-field (BF) images, along with EDS analyses of both the Cu/Mo and Mo/Mg₃Bi_1.5Sb_0.5 interfaces, further confirm these observations (Supplementary Figs. 6 and 7). Furthermore, as shown in the Cu-Mo binary phase diagram (Supplementary Fig. 8), no intermetallic compounds are formed between Cu and Mo. Together with the Mg-Bi-Mo ternary phase diagram, this suggests that Mo can act as an effective diffusion barrier at high temperatures, improving the service stability of the interface. High-resolution transmission electron microscopy (TEM) reveals that the Mo/Mg₃Bi_1.5Sb_0.5 interface is largely disordered and rough, with pronounced phase-contrast variations (Fig. 2c), indicative of incoherent characteristics. The rough interfacial morphology and phase-contrast variations suggest local structural disorder generated during heterogeneous atomic-scale bonding between the sputtered Mo layer and the Mg₃Bi_1.5Sb_0.5 substrate. Owing to the differences in crystal structure, bonding characteristics, and atomic arrangement between Mo and Mg₃Bi_1.5Sb_0.5, incoherent interfacial regions may form during sputter deposition, leading to the observed contrast fluctuations. These features indicate effective Mo/Mg₃Bi_1.5Sb_0.5 interfacial bonding and support the feasibility of magnetron sputtering for constructing Mg-based thermoelectric interfaces. Fast Fourier transform (FFT) analysis shows that Mo retains a cubic $\mathrm{Im}\overline{3}\mathrm{m}$ structure (Fig. 2d, e), whereas Mg₃Bi_1.5Sb_0.5 preserves its trigonal $\mathrm{P}\overline{3}\mathrm{m}1$ structure (Fig. 2f, g).

Fig. 2. Microstructural characterization of n-type junction.

a HAADF-STEM image of the as-prepared Mg_3.2Bi_1.4975Sb_0.5Te_0.0025/Mo/Cu junction with corresponding EDS elemental mapping. b EDS elemental line scan along the arrow in (a). c–g High-resolution TEM images of the as-prepared Mg_3.2Bi_1.4975Sb_0.5Te_0.0025/Mo junction (c), along with enlarged views of Mo (d) and Mg₃Bi_1.5Sb_0.5 (f), and their corresponding FFT patterns: Mo (e) and Mg₃Bi_1.5Sb_0.5 (g).

The Ag/Mo/MgAgSb interface, after mechanical polishing, exhibits sharply delineated layers with strong interfacial bonding. The inclusion of Mo improves adhesion within the junction, as reflected by the robustly bonded architecture observed post-polishing (Supplementary Fig. 9). The Ag surface presents exceptional smoothness, density, and uniformity (Supplementary Fig. 10). HAADF-STEM imaging, along with corresponding elemental maps (Supplementary Fig. 11), reveals uniform elemental distribution of Ag, Mo, and MgAgSb, with no sign of segregation. The thicknesses of the Mo and Ag layers are ~0.5 μm and ~2 μm, respectively. High-resolution TEM (Supplementary Fig. 12) identifies a disordered atomic configuration with distinct phase contrast near the MgAgSb/Mo interface, indicative of incoherent characteristics. FFT results indicate that Mo crystallizes in the cubic $\mathrm{Im}\overline{3}\mathrm{m}$ phase, while MgAgSb retains its tetragonal $\mathrm{I}\overline{4}2\mathrm{c}$ structure (Supplementary Fig. 12).

Electrical contact resistivity and thermal stress

As shown in Fig. 3a, the Cu/Mo/Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 junction exhibits a low contact resistivity (ρ_c) of ~5.9 μΩ cm², measured via the four-probe method. To assess the thermal stability of the sputtered interface, three Cu/Mo/Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 junction samples were prepared using the same deposition process and subjected to different aging conditions: unannealed, vacuum-annealed at 250 °C for 71 h, and vacuum-annealed at 250 °C for 166 h. As shown in Supplementary Fig. 13, the interface remained sharp and continuous after aging, with no detectable cracks or interfacial gaps. Moreover, the contact resistivity also showed negligible change after annealing, further confirming the excellent bonding integrity and thermal stability of the junction below 300 °C. In contrast, the Ag/Mo/MgAgSb junction shows a higher contact resistivity of 51.7 μΩ cm², indicating that further optimization is required (Supplementary Fig. 14). The ratio of contact to total resistance (R_c/R_total) of our n-type leg was derived as a function of leg length and contact resistivity, assuming a bulk resistivity of ~14 μΩ m (Fig. 3b), as the thermoelectric leg length is greatly reduced. The contact resistance contributes ~15.3%, 4.3%, and 2.9% at leg lengths of 0.5, 2, and 3 mm, respectively. At a leg length of 0.5 mm, the Cu/Mo/Mg₃Bi_1.5Sb_0.5 junction exhibits a 44.9% lower contact resistance fraction than the two-step hot-pressed Cu/Mg₂Cu/Mg₃Bi_1.5Sb_0.5 counterpart^22,44. Mechanical testing revealed tensile strengths of 3.9 MPa and 4.1 MPa for the Cu/Mo/Mg₃Bi_1.5Sb_0.5 and Ag/Mo/MgAgSb junctions, respectively, both with a cross-sectional area of 3.5 × 4.1 mm² (Supplementary Fig. 15). Owing to their robust bonding strength, Mg₃Bi_1.5Sb_0.5 and MgAgSb thermoelectric legs with integrated interface layers (0.48 × 0.48 mm²) were precisely cut and reliably soldered to Cu electrodes on an AlN ceramic substrate (Fig. 3c).

Fig. 3. Interfacial contact performance of thermoelectric junctions.

a Scanning resistance (R) as a function of distance of Cu/Mo/Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 junction with an area of 3.68 × 3.2 mm², showing an contact resistivity of 5.9 μΩ cm². b Variation of the contact resistance proportion in n-type thermoelectric legs with respect to leg length (h) and contact resistivity (ρ_c), assuming a bulk resistivity of 14 μΩ m. c Photographs of the cut n-type and p-type thermoelectric legs soldered onto Cu electrodes on AlN ceramic plates. d Schematic diagram of the μ-TEC structure, including thermoelectric legs, interface layers (Cu/Mo and Ag/Mo), solder layers, electrode layers, and ceramic substrates. e Simulated average thermal stress without the Mo interlayer for interface layer thicknesses of 3 μm and 150 μm (cases 1 and 2, respectively). f Simulated average thermal stress for Cu/Mo and Ag/Mo interfaces with thicknesses of 3 μm/1 μm and 112.5 μm/37.5 μm (cases 3 and 4, respectively).

To study the effect of the Mo layer on device thermal stress, we performed finite element simulations of four interface configurations under constant material and boundary conditions (Supplementary Table 1). The fourth case corresponds to the experimentally fabricated μ-TECs, with the structure illustrated in Fig. 3d and parameters detailed in Supplementary Table 2. A thermal load corresponding to the device’s maximum cooling power at ΔT = 0 K was applied at the cold end to simulate actual operating conditions. Given that the total height of the thermoelectric leg was fixed, the increased thickness of the interface layers in Case 1 and Case 3 effectively reduced the height of the thermoelectric material by ~50%, thus halving the device’s cooling power. Accordingly, a maximum heat load of 5 W cm⁻² was assigned to Cases 1 and 3, while 10 W cm⁻² was applied to Cases 2 and 4. In all cases, the applied maximum heat load at the cold side resulted in a peak cooling temperature difference below 1 K (Supplementary Fig. 16). Consequently, the small temperature gradients in the device gave rise to uniform and low thermal stress levels (<10 MPa), as illustrated in Supplementary Fig. 16. To evaluate stress within the thermoelectric legs, multiple cross-sections were extracted for all four cases (Supplementary Figs. 17–20). In all cases, the p-type legs exhibited higher stress than the n-type legs. Interfacial stress increased markedly as interface layer thickness decreased, regardless of Mo presence, with pronounced stress concentrations at edges and corners. Axial stress profiles were quantified by tracking the maximum and minimum stresses along the leg axis. As shown in Supplementary Fig. 21, maximum thermal stress is not determined solely by the coefficient of thermal expansion; reducing the interface layer thickness markedly increases stress in both the interface and thermoelectric materials (Case 1 vs. 2, Case 3 vs. 4). When the interface layer is comparable in thickness to the thermoelectric leg, Mo raises maximum stress in both the interface and thermoelectric materials (Case 1 vs. 3). For very thin interfaces, Mo increases stress in the interface while slightly reducing it in the thermoelectric material (Case 2 vs. 4). As shown in Supplementary Fig. 22, minimum thermal stress in both the interface and thermoelectric material decreases with reduced interface thickness (Case 1 vs. 2, Case 3 vs. 4). Introducing a Mo layer raises the minimum stress in the interface but simultaneously lowers the minimum stress in the thermoelectric material (Case 1 vs. 3, Case 2 vs. 4). As shown in Fig. 3e, thinning the interface layer without altering its composition has a relatively minor impact on the average thermal stress along the thermoelectric leg (Case 1 vs. 2). With the addition of the Mo layer, the average thermal stress increases from 4.5 to 14.5 MPa in the interface layer and from 3.5 to 4.0 MPa in the thermoelectric layer (Case 1 vs. 3, Fig. 3e, f). With a substantial reduction in interface thickness (Case 3 vs. 4), the average stress decreases to 12.7 MPa in the interface layer and 3.0 MPa in the thermoelectric layer, accompanied by a more uniform stress distribution across the latter (Fig. 3f). In addition, thermal stress distributions were simulated for the four cases under a temperature difference of ~32 K (Supplementary Figs. 23–28). Compared with the maximum-cooling-power condition, a larger temperature gradient raised the maximum stress to around 300 MPa, while the average stress remained below 200 MPa. Thermal stress was primarily concentrated at the edges and corners, and the overall magnitude still stayed lower than the typical stress levels if conventional TEG modules, which can exceed 600 MPa⁵¹. Overall, thermal stress is governed not only by thermal expansion mismatch but also by interface layer thickness, device geometry, and the mechanical properties of thermoelectric materials. Therefore, although the Mo layer elevates local interfacial stress, it does not increase stress within the thermoelectric legs, nor does it counteract the performance advantages brought by optimized interfacial transport and diffusion barrier effects.

Cooling performance of μ-TECs

Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 was synthesized via our previously developed method²⁹, achieving a room-temperature zT of ~0.92 (Fig. 4a) and a power factor of ~33 μW cm⁻¹ K⁻² (Supplementary Fig. 29), owing to a high carrier mobility (~206 cm²V⁻¹ s⁻¹, Supplementary Table 3). This is slightly below the single parabolic band model (SPB) prediction, likely due to neglect of grain boundary scattering^29,44 (Supplementary Fig. 30). Phase-pure α-MgAgSb was prepared by two-step ball milling and induction sintering (Supplementary Fig. 31), reaching a room-temperature zT of ~0.95, a power factor of ~24 μW cm⁻¹ K⁻², and a Hall mobility of 97 cm²V⁻¹ s⁻¹ (Supplementary Table 3). Notably, our Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 exhibits a compressive strength of 722 MPa (Supplementary Fig. 32), which, along with MgAgSb, surpasses commercial zone-melted Bi₂Te₃ by 28.9- and 15.6-fold, and is 2–4 times higher than hot-extruded Bi₂Te₃, typically used in Bi₂Te₃-based μ-TECs (Supplementary Fig. 33).

Fig. 4. zT values and cooling performance of Mg-based μ-TEC.

a Temperature-dependent zT values of n-type Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 and p-type MgAgSb. b, c Calculated maximum temperature difference (ΔT_max) (b) and maximum cooling power density (q_cmax) (c) as functions of the area ratio (A_n/A_p) for thermoelectric leg lengths of 0.2, 0.5, 1.0, 2.0, 3.0, and 10.0 mm, with the p-type leg cross-sectional area fixed at 0.25 mm². d Calculated and measured ΔT (hot side at 300 K and 325 K), COP (hot side at 353 K, ΔT = 40 K), and q_c (hot side at 300 K) as functions of current (I). e Calculated q_cmax based on the thermoelectric and interface properties reported herein, as a function of thermoelectric leg length (h) and fill factor, alongside comparison with reported experimental data for Mg₃(Bi,Sb)₂- and MgAgSb-based cooling devices^22,44–46. f Calculated q_cmax as a function of the ceramic plate thickness (h_c) and the contact resistivity of the p-type thermoelectric leg (ρ_c).

Prior to fabrication, calculations using the parameters in Supplementary Table 4 were conducted to determine the effect of leg dimensions on maximum temperature difference (ΔT_max), cooling power density (q_c), and coefficient of performance (COP). Short leg lengths in μ-TECs make contact resistance a significant contributor to total device resistance, strongly degrading performance. Consequently, the maximum cooling temperature difference (ΔTₘₐₓ) was calculated as a function of the n/p leg area ratio (A_n/A_p) for varying leg lengths (Fig. 4b). The optimal A_n/A_p increases with leg length, ranging from 0.43 to 0.75 as the leg height varies from 0.2 mm to 10 mm, with Aₚ fixed at 0.25 mm². For a 0.5 mm leg, the optimal ratio is ~0.53, corresponding to an n-type leg area of 0.36 × 0.36 mm². For all leg lengths, q_cmax grows with A_n/A_p (Fig. 4c), and shorter legs amplify this effect. Specifically, for a thermoelectric leg length of 0.5 mm, increasing A_n/A_p from 0.53 to 1.0 reduces ΔTₘₐₓ by 6.0% while raising q_cmax by 24.6%, demonstrating that a larger n/p area ratio enhances q_cmax. To prevent short-circuiting during soldering, μ-TEC thermoelectric legs were designed with a 0.48 mm cross-section, just below the Cu electrode width (~0.55 mm), achieving a high fill factor of 43%. A schematic of Mg-based μ-TEC fabrication is shown in Supplementary Fig. 34. The high adhesion of the Mo layer enabled successful interfacial layer deposition (Supplementary Fig. 35) and precise cutting of thermoelectric legs (Supplementary Fig. 36). Figure 4d shows measured ΔT, q_c, and COP versus current (I). At I = 0.8 A, ΔT reached 32 K and 42 K for hot-side temperatures of 300 K and 325 K. At I = 0.43 A, with a hot-side temperature of 353 K and a ΔT of 40 K, the device achieved a COP of 0.24. Notably, q_cmax measured at 300 K reached 4.34 W cm⁻², slightly below the simulated values. The difference between the experimental and theoretical values is likely due to additional resistance introduced during soldering. As shown in Supplementary Table 5, the resistance contributions from different components of the device were calculated. The discrepancy likely arises from soldering-induced resistance and the unaccounted contribution of the 25-μm-thick Cu electrodes, leading to a measured resistance of 1.78 Ω, slightly above the theoretical value of 1.56 Ω. Despite this deviation, the device exhibits good reproducibility, with coefficients of variation (CVs) of 1.38% for resistance and 1.78% for ΔT_max across multiple measurements (Supplementary Table 6). We further evaluated q_cmax as a function of leg length and fill factor and compared with reported Mg-based devices^22,44–46. In most prior MgAgSb- or Mg₃(Bi,Sb)₂-based devices, leg length exceeds 2 mm, and fill factor is <25%, limiting q_cmax to ~1 W cm⁻². Our device achieves 4.3 W cm⁻²—3–4 times higher—at a leg height of 0.56 mm and fill factor of 43% (Fig. 4e). In addition, simulations of q_cmax as a function of ceramic plate thickness (h_c) and p-type contact resistivity (Fig. 4f) indicate that lowering the resistivity to 5 μΩ cm² and h_c to 0.2 mm could raise q_cmax to 9.5 W cm⁻². Further optimization of soldering, reduced contact resistivity of MgAgSb interface layers, and higher power factor could increase the maximum cooling power density of Mg-based μ-TECs, potentially reaching 20 W cm⁻². Owing to the lack of experimental setups for power generation, performance was evaluated by calculation (Supplementary Fig. 37). At ΔT = 150 K, the device delivers a maximum power (P_max) of ~0.095 W, corresponding to a power density (ω_max) of 0.73 W cm⁻² and an efficiency of ~5%.

In summary, using a low-temperature and water-free fabrication route, we produced Cu/Mo/Mg₃(Bi,Sb)₂ and Ag/Mo/MgAgSb junctions with high bonding strength, achieving 12-pair MgAgSb/Mg₃(Bi,Sb)₂ μ-TECs with a packing density of 93.5 leg pairs per cm², a cold-side area of 2.95 × 4.35 mm², a fill factor of 43%, and a cooling power density of 4.34 W cm⁻², substantially surpassing previous reports on Mg-based thermoelectric devices. These results establish a scalable, low-temperature and water-free route to fully exploit emerging thermoelectric materials in micro-cooling and thermal management. Looking forward, overcoming bonding challenges is essential for high-quality interfaces and the widespread deployment of reliable, high-performance micro-thermoelectric modules based on emerging materials.

Methods

Material syntheses

Mg turnings (Western Elements, 99.99%), Sb shots (Aladdin, 99.99%), Bi shots (Aladdin, 99.99%), Ag shots (Macklin, 99.99%), and Te shots (Aladdin, 99.99%) were weighed according to the nominal stoichiometry of Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 and MgAgSb. Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 powder was prepared by high-energy ball milling for 12 hours under an argon atmosphere using a SPEX 8000D Mixer/Mill (SPEX, USA). Subsequently, pellets of Mg_3.2Bi_1.4975Sb_0.5Te_0.0025, approximately 12.7 mm in diameter and 11 mm thick, were prepared by hot pressing at 1053 K for 10 minutes. A two-step milling process was employed for MgAgSb. Mg and Ag were ball-milled at an atomic ratio of 1:1 for 10.5 h under an Ar atmosphere using a SPEX 8000D Mixer/Mill (SPEX, USA). Subsequently, Sb was added to achieve a final stoichiometry of Mg:Ag:Sb = 1:1:1, followed by an additional 10.5 h of milling. MgAgSb pellets with a diameter of 12.7 mm and a thickness of approximately 9 mm were hot pressed at 553 K for 1 hour.

Interface layer preparation

The sintered ingots, 12.7 mm in diameter, were cut into 0.56 mm-thick slices using wire electrical discharge machining. For the Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 samples, Cu/ Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 junctions were prepared by directly sputtering Cu at 60 W for 80 min under a pressure of 0.4 Pa, while Cu/Mo/Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 junctions were fabricated by first depositing Mo at 50 W or 70 W for 40 min under 0.4 Pa, followed by Cu deposition at 60 W for 80 min under the same pressure. For the MgAgSb samples, Ag/MgAgSb junctions were prepared by directly sputtering Ag at 60 W for 1 h under 0.4 Pa, while Ag/Mo/MgAgSb junctions were fabricated by first depositing Mo at 45 W for 40 min or 65 W for 35 min under 0.4 Pa, followed by Ag deposition at 60 W for 1 h under the same pressure.

Fabrication of micro-thermoelectric modules

The thermoelectric slices with deposited interface layers, 0.56 mm thick, were diced into square blocks measuring 0.48 × 0.48 mm² using wire electrical discharge machining. Subsequently, the n-type Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 and p-type MgAgSb thermoelectric legs were soldered onto a 0.4 mm-thick aluminum nitride (AlN) ceramic substrate using Sn63/Pb37 solder. The AlN ceramic substrate (2.95 ×  4.35 mm²) featured copper electrodes (0.55 × 1.25 mm²) with an inter-electrode gap of 0.15 mm.

Characterization and measurement

Powder X-ray diffraction (XRD) data were collected using a Bruker D8 Advance diffractometer equipped with a Cu Kα radiation source (λ = 1.5406 Å). Electrical resistivity and Seebeck coefficient measurements were performed using a ZEM-3 system (ULVAC-RIKO, Japan). Thermal conductivity ($\kappa$) was calculated via the relation ${\kappa =\rho DC}_{\mathrm{P}}$, where the mass density ($\rho$) was measured by the drainage method, the specific heat capacity at constant pressure ($C_{\mathrm{P}}$) was estimated based on the Dulong–Petit law, and the thermal diffusivity ($D$) was measured using a LFA 467 laser flash apparatus (NETZSCH, Germany). The microstructure and chemical composition of the samples were analyzed using a Zeiss Sigma 300 thermal field emission scanning electron microscope (SEM) equipped with energy-dispersive X-ray spectroscopy (EDS). TEM samples were prepared using a focused ion beam (FIB) system (FEI Scios 2 HiVac). Conventional transmission electron microscopy (TEM) images were acquired using an FEI Talos F200X G2 microscope. The room-temperature Hall carrier concentration and Hall mobility were measured using a Toho Technology HL9900 Hall system. The compressive strength of Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 and the strength of the junctions were determined using a universal testing machine, model CMT4503. Cu/Mo/ Mg_3.2Bi_1.4975Sb_0.5Te_0.0025 and Ag/Mo/MgAgSb junction samples with cross-sectional areas of 3.68 by 3.2 mm and 4.52 by 3.27 mm were tested for contact resistance via the four-probe method. Total contact resistivity was calculated using ${\rho }_{\mathrm{c}}={AR}_{jump}$. The mechanical properties were measured using a CMT4503 universal testing machine. The cooling performance of our Mg-based μ-TEC was measured by Ferrotec Co., Ltd., Hangzhou, China. All cooling performance tests, including temperature difference (ΔT), cooling power (Q_c), and coefficient of performance (COP), were conducted under vacuum conditions.

Temperature difference measurement: The hot side of the device was mounted on a temperature-controlled stage and maintained at ~25 °C, while the cold side was thermally insulated to achieve near-adiabatic conditions. An optimized current was applied, and ΔT was measured using thermocouples positioned on both sides of the device.

Q_c and COP measurement: The hot side was fixed at ~25 °C. A micro-heater attached to the cold side provided a heat load assumed to be fully transferred to the device (Q_c = P_heater), neglecting heat losses. The heater power was adjusted until ΔT = 0, at which point Q_c was obtained.

The cooling power density (q_c) was calculated as q_c = Q_c / A, where A is the cold-side area. The COP was calculated as COP = Q_c / P, where P is the electrical input power.

Numerical calculation and simulation

The temperature difference (∆T), cooling capacity (Q_c), and coefficient of performance (COP) were calculated using the following equation, which is based on the thermal balance condition among Joule heating, thermal conduction, and the Peltier effect.

$$S{T}_{\mathrm{c}}I-\frac{1}{2}{I}^{2}R-K△T=0$$ 1

$$Q_{\mathrm{c}}=S{T}_{\mathrm{c}}I-\frac{1}{2}{I}^{2}R-K△T$$ 2

$$COP=\frac{S{T}_{\mathrm{C}}I-\frac{R}{2}{I}^2-K\mathrm{\Delta }T}{I^{2}R+SI\mathrm{\Delta }T}$$ 3

where S is the total Seebeck voltage of the device, R is the total electrical resistance of the thermoelectric elements, K denotes the thermal conductance, I is the electric current, T_c represents the temperature at the cold end, and ΔT is the temperature difference between the hot and cold sides.

In addition, the power generation performance is calculated based on the following formula:

$$I=\frac{S△T}{R+R_{\mathrm{l}}}$$ 4

$$P=\frac{S^2{△T}^2{R}_{\mathrm{l}}}{{\left(R+R_{\mathrm{l}}\right)}^2}$$ 5

$$Q_{\mathrm{h}}=S{T}_{\mathrm{h}}I-\frac{1}{2}{I}^{2}R+K△T$$ 6

$$\eta =\frac{P}{Q_{\mathrm{h}}}$$ 7

Where $R_{\mathrm{l}}$ represents the load resistance, $Q_{\mathrm{h}}$ the heat flux at the hot side, $P$ the output power, and $\eta$ the power generation efficiency.

Due to the small geometric dimensions of the device, contact resistance and thermal resistance exert a considerable influence on its performance. Consequently, the cooling performance calculations must account for the thermal conductivity of the ceramic substrate as well as the contact resistance. Following calibration, the device resistance and the actual temperature difference across the ceramic plates are determined by the equation below.

$$R_{\mathrm{total}}=R+2\frac{{\rho }_{\mathrm{c},\mathrm{n}}}{A_{\mathrm{n}}}+2\frac{{\rho }_{\mathrm{c},\mathrm{p}}}{A_{\mathrm{p}}}$$ 8

$${\mathrm{\Delta }T}^{\prime }=△T\left(1+2\frac{N{\kappa }_{\mathrm{c}}}{h_{\mathrm{c}}}\frac{1}{K}\left(A_{\mathrm{N}}+A_{\mathrm{P}}\right)\right)-2\frac{S\overline{T}I{h}_{\mathrm{c}}}{N\left(A_{\mathrm{N}}+A_{\mathrm{P}}\right){\kappa }_c}$$ 9

where ${\rho }_{\mathrm{c},\mathrm{n}}$ and ${\rho }_{\mathrm{c},\mathrm{p}}$ are the contact resistivities of the n-type and p-type thermoelectric leg interfaces, respectively; $A_{\mathrm{N}}$ and $A_{\mathrm{P}}$ denote the cross-sectional areas of the n-type and p-type legs; $h_{\mathrm{c}}$ is the thickness of the ceramic plate; ${\kappa }_{\mathrm{c}}$ represents the thermal conductivity of the ceramic plate; $N$ is the number of thermoelectric leg pairs; $\overline{T}$ is the average temperature; $R$ is the total resistance of the thermoelectric materials; $R_{\mathrm{total}}$ is the total device resistance; and $△T$, ${\mathrm{\Delta }T}^{\prime }$ refer to the temperature difference across the thermoelectric materials and ceramic plate, respectively.

The mobility and Seebeck coefficient are calculated using the following equations:

$$\frac{1}{\mu }=\frac{1}{{\mu }_{\mathrm{ph}}}+\frac{1}{{\mu }_{\mathrm{al}}}$$ 10

$${\mu }_{\mathrm{ph}}=\frac{\sqrt{2}\mathrm{\pi }e{ħ}^4}{3{\left(k_{\mathrm{B}}T\right)}^{\frac{3}{2}}}\frac{v_{\mathrm{l}}^{2}d}{E_{\mathrm{def}}^2{\left(m_{\mathrm{b}}^{*}\right)}^{\frac{3}{2}}{m}_{\mathrm{I}}^{*}}\frac{F_0\left(\eta \right)}{F_{\frac{1}{2}}\left(\eta \right)}$$ 11

$${\mu }_{\mathrm{al}}=\frac{16e{ħ}^4}{9{\sqrt{2}\mathrm{\pi }x\left(1-x\right)\left(k_{\mathrm{B}}T\right)}^{\frac{1}{2}}}\frac{N_0}{E_{\mathrm{al}}^2{\left(m_{\mathrm{b}}^{*}\right)}^{\frac{3}{2}}{m}_{\mathrm{I}}^{*}}\frac{F_0\left(\eta \right)}{F_{\frac{1}{2}}\left(\eta \right)}$$ 12

$$S=-\frac{k_{\mathrm{B}}}{e}\left(\frac{\left(2+\lambda \right)F_{\mathrm{\lambda }+1}\left(\eta \right)}{\left(1+\lambda \right)F_{\mathrm{\lambda }}\left(\eta \right)}-\eta \right)$$ 13

$$F_{\mathrm{i}}\left(\eta \right)={\int }_0^{\infty }\frac{x^{i}dx}{1+e^{x-\eta }}$$ 14

Where $v_{\mathrm{l}}$ is the longitudinal sound velocity, $d$ the density, $N_0$ the number of atoms per unit volume, $x$ the fractional concentration, and $E_{\mathrm{def}}$ and $E_{\mathrm{al}}$ are the deformation and alloy scattering potentials.

The temperature and thermal stress distributions in the thermoelectric device were simulated using finite element calculations, with the relevant thermoelectric parameters derived from experimental measurements, as shown in the Supplementary Fig. 29. The calculation did not consider interface resistivity. The parameters used for the thermal stress simulation are listed in Supplementary Table 2. The phase diagram calculations were based on data obtained from the Materials Project website (https://next-gen.materialsproject.org/materials).

Author contributions

L.-M. Wu, L. Chen, and J. Yang conceived the idea. J. Yang synthesized n-type Mg₃Bi_1.5Sb_0.5 and p-type MgAgSb samples, characterized their thermoelectric properties, and performed SEM and TEM sample preparation. R. Zhu fabricated the Cu/Mo/Mg₃Bi_1.5Sb_0.5 and Ag/Mo/MgAgSb junctions by magnetron sputtering. J. Yang fabricated Mg-based μ-TECs and carried out detailed numerical simulations and finite-element analyses. M. Li measured the performance of the Mg-based μ-TECs. J. Yang, Z. Mei, L. Chen, and L.-M. Wu wrote and edited the manuscript. All authors contributed to helpful discussions.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This research was supported by the National Key Research and Development Program of China (no. 2022YFB3803902).

Data availability

The data supporting the findings of this study are available within the Article and its Supplementary Information. Source data are provided with this paper.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-75165-8.