Design and realization of high performance textured lead-free piezoelectric ceramics through human-AI collaboration Authors: Aman Nanda, Debjyoti Bhattacharya, Nayeon Kang, Shankar Kunwar, Shashank Priya, Jungho Ryu, Wesley Reinhart, Michael Lanagan, Bed Poudel Published: 2026 ABSTRACT Developing multielement doped Pb-free (K,Na)NbO₃ piezoelectrics often hindered by complex doping trends and tedious trial-and-error experimentation. Here, we present a human-in-the-loop, artificial intelligence guided materials design framework that utilizes large language models to capture implicit structure-property knowledge from prior literatures and propose new compositions. Expert intervention further directs experimental realization based on materials science principles and experiential knowledge, accelerating discovery of targeted compositions. Using collaborative strategy, synthesized random composition exhibiting piezoelectric charge constant d33 of 440 − 500 pC/N which further enhanced to 600-620 pC/N through crystallographic texturing and sintering aid optimization. Despite inherently off-MPB degradation (at R.T), this composition maintained steady electromechanical coupling (kij) and d31 up to 160 oC. To validate practical relevance, a cantilever-based magneto-mechano-electric (MME) energy harvester was fabricated, delivering a power density of ~ 705μW/cm3 at the second harmonic, outperforming reported Pb-free MME designs. Here, we demonstrate an exceptional approach towards developing application-specific functional materials through the synergy of AI-driven recommender systems, human expert validation, and experimental realization. Introduction Pb-free piezoelectric material, (K, Na)NbO3 composition is an attractive candidate with a low theoretical density of 4.5 g/cm3 and the ability to form complete solid solutions with alkali, alkaline, and rare-earth metals, as well as transition metals1. The issue lies with inherently low electromechanical properties (d33 ~ 160 pC/N) in base composition, even after meticulous optimization of the K/Na ratio2. To address the issue, the co-doped strategy has been profoundly used, where the base K0.5Na0.5NbO3 (KNN) composition is modified with more than two dopants, each contributing to a particular set of properties. Apart from individual elements, in recent years, the doping methodology has evolved in form of additional ABO3 systems like Bi0.5A0.5ZrO3 (A: Na+, K+, Ag+, and Li+ etc.) and BiBO3 (B: Al3+, Fe3+, and Sc3+) which consolidates both the ferroelectric-ferroelectric phase boundary (R-O and O-T) at room temperature resulting in significant improvement of electrical properties along with diffused transition resulting in temperature stability3,4. By developing microstructural orientation (texturing), the magnitudes can be further enhanced by exploiting the crystal anisotropy that exhibit the most favorable properties. Literature reports, textured R-O-T multiphase KNN compositions proved to be a benchmark against conventional MPB-type systems (such as the PZT system) in terms of figure of merit (FoM) required for tonpilz transducer5. Broadly, the FoM of piezoelectric materials can be tailored for specific applications. High-power devices, such as transducers, transformers, and ultrasonic motors, require a high mechanical quality factor (Qm) with stable electromechanical coupling (kij) and piezoelectric coefficient (dij) to operate reliably at resonance6. However, resonant operation introduces challenges including self-heating, domain-wall-induced nonlinearity, and thermal runaway, potentially leading to depolarization7,8. These effects are mitigated through acceptor doping, which produces “hard” compositions suitable for high powered operation9,10. In contrast, low-power or off-resonance applications, such as precision sensors, actuators, and energy harvesting, demand high sensitivity, achieved via enhanced dij and voltage coefficients (gij) in donor-doped “soft” compositions. In both regimes, thermal stability is critical to maintain linearity and prevent depolarization. While these application-specific performance requirements are well understood, achieving them simultaneously through compositional engineering of complex piezoelectric systems remains a significant materials design challenge. The optimization of multi-element compositions is typically governed by tedious, iterative experimentation, with limited systematic frameworks available to disentangle the individual and synergistic roles of multiple dopants. This issue can be addressed using complex pattern recognition techniques from the existing literature, employing a data-driven approach. Additionally, generative modeling approaches, especially large language models (LLMs), has advanced pattern recognition capabilities and offer a promising route to uncover the correlation between complex dopants and their corresponding properties11,12. LLMs, despite originally being developed for text generation, have recently emerged as powerful tools for pattern recognition, knowledge retrieval, and hypothesis generation in scientific and engineering contexts13–17. We hypothesized that LLMs can help guide human experts through high-dimensional design spaces by identifying correlations, inferring dopant trends, and proposing composition-phase relationships. Although these models are effective at extracting statistical patterns across large textual corpora, their ability to perform causal reasoning remains unproven. To mitigate this problem, a human-in-the-loop framework is adopted where the role of the LLM is explicitly constrained to the generation of chemically diverse candidate compositions within expert-defined design spaces. This division of labor is intentionally structured to utilize the complementary strengths of both LLM (facilitating efficient exploration of a high-dimensional compositional landscape) and human intervention (ensuring adherence to physically meaningful principles and guarding against heuristic biases). Utilizing the pattern recognition capabilities of LLMs along with expert refinement, this work accelerates the discovery of new Pb-free compositions for high-performance piezoelectric materials. The schematic of the work is shown in Fig. 1. Firstly, a few-shot prompting approach was adopted to guide the model (Claude 3.5 Sonnet) for interpreting complex literature reports for a meaningful pattern. Furthermore, the model was designed to meet application-specific requirements, yielding a high d33 value with a multiphase composition (R-O-T) at room temperature. Through multiple iterations, the outcome was generated via a dual segregation approach to broadly categorize the compositions based on their predicted magnitude of d33. Before validation in the laboratory, the compositions were further filtered based on charge balance and dopant stoichiometry. One composition from each group was synthesized and characterized to verify the model’s predictive capability. Furthermore, a post-hoc comparative analysis was conducted to evaluate the predictive accuracy, generalization ability, adaptive reasoning, and few-shot learning behavior of d33 across multiple LLM models.Fig. 1Schematic illustration of the LLM-assisted compositional discovery framework for KNN-based piezoelectric ceramics.a Few-shot prompting with literature-derived multiphase compositions guides the generation of chemically valid candidates under stoichiometric and phase-fraction constraints. Targeted phase engineering enables the identification of high-performance compositions by explicitly optimizing crystallographic phase coexistence. b Manual refinement (human-in-the-loop approach) to filter the synthesizable composition. Predictions from the model are purely statistical, and in some cases violate the stoichiometry or the charge balance between A/B sites, which requires a mandatory physics-based sanity check. c Post-hoc property prediction analysis result of the best composition across different LLM models, which defines the predictive capability and the variance across a broad spectrum of A.I. models, d Characterization of desired composition for superior electrical properties along with its texturing. Appropriate microscopy to validate the texture degree and crystallographic orientation. Textured ceramics are supposed to exhibit superior properties due to crystal anisotropy; its outcome is validated in the form of multiple electrical measurements. Schematic illustration of the LLM-assisted compositional discovery framework for KNN-based piezoelectric ceramics. a Few-shot prompting with literature-derived multiphase compositions guides the generation of chemically valid candidates under stoichiometric and phase-fraction constraints. Targeted phase engineering enables the identification of high-performance compositions by explicitly optimizing crystallographic phase coexistence. b Manual refinement (human-in-the-loop approach) to filter the synthesizable composition. Predictions from the model are purely statistical, and in some cases violate the stoichiometry or the charge balance between A/B sites, which requires a mandatory physics-based sanity check. c Post-hoc property prediction analysis result of the best composition across different LLM models, which defines the predictive capability and the variance across a broad spectrum of A.I. models, d Characterization of desired composition for superior electrical properties along with its texturing. Appropriate microscopy to validate the texture degree and crystallographic orientation. Textured ceramics are supposed to exhibit superior properties due to crystal anisotropy; its outcome is validated in the form of multiple electrical measurements. As strategized, the predicted random composition 0.96(K0.5Na0.5)(Nb0.94Sb0.06)O3-(0.03)(Bi0.5Na0.4K0.1)(Zr0.96Ti0.04)O3-0.01(Sr0.7Ca0.3)ZrO3 – x mol % Fe2O3 (x = 0, 0.1, 0.2) (KNN-BNKZT-SCZ) is successfully synthesized with a d33 ~ 440–500 pC/N which is approximately 3 times higher than the pristine KNN composition2. The phase fraction attained was in conjecture with the prompted ratios to the model. The amount of Fe2O3 affected the final phase fraction (calculated through Rietveld refinement), consequently, the electrical properties. The developed composition was further textured and optimized with 3 wt% NaNbO3 templates along with a 0.1 mol% Fe2O3 sintering aid to achieve a d33 of 620 pC/N. With texturing, the magnitude of d33 and εr behaved conversely, improving the g33 to fulfill the requirement in ME and other off-resonance applications. To validate the performance advantage of LLM-predicted and experimentally synthesized composition, MME (magneto-mechano-electric) response of the cantilever structured composites with both random and textured materials was evaluated, including the MME energy harvesting capabilities. MME generators, which typically use piezoelectric-magnetostrictive composites, also termed magnetoelectric (ME) composites, are gaining importance due to their ability to facilitate autonomous power supply for structural health monitoring systems, wireless power transfer, and low-frequency wireless communication technology18,19. Recent progress demonstrates that MME technology can provide sufficient standalone power for small signal Internet of Things sensors, including the WiFi20,21. A significant difference in αDME of random and textured compositions was observed and measured up to 27 mV/cm.Oe and 90 mV/cm.Oe respectively. Further, the maximum RMS power density in textured composition was measured up to 705 μW/cm3, and that of random was 1 μW/cm3. This study establishes LLMs as an effective data-driven paradigm for uncovering hidden compositional-phase-property relationships, enabling accelerated discovery of high-performance, Pb-free piezoelectric materials with application-relevant functionality. Results and discussion Analysis of predicted compositions As aforementioned, Fig. 1 presents the steps involving the discovery of KNN-based piezoelectric ceramics driven by structured LLM-assisted compositional discovery. The efficacy of the seven generative strategies described in Table S1 varied significantly. While strategies focusing on simple elemental substitution (1–3) yielded chemically valid but conventional candidates, Strategy 7 (targeted phase engineering) proved superior. By prompting the model to optimize for explicit phase coexistence fractions, the generator successfully identified the lead candidate C1 (d33 ∼ 660 pC/N). This suggests that injecting explicit, expert-provided microstructural knowledge into the zero-shot instructions aligns the stochastic token generation more closely with physically realizable materials than the injection of composition alone. To maintain parity between reported compositions, derived inference, and generative synthesis, ten multiphase reported compositions, parameterized in the general function (1-x-y-z-w)(K, Na)NbO3 – xA1B1O3 – yA2B2O3 – zA3B3O3 – wA4B4O3), were supplied to the model through few-shot prompting as a chemically coherent anchoring (Table S2). Existing literature consistently reports that doping above 0.05 molar fraction or (1-x-y-z-w) < 0.95 limits grain growth, which deteriorates long-range ferroelectric order and suppresses functional coefficients, including Pr, dij, and kij4,22. Based on this empirical estimation, prompting constraints were applied such that (1-x-y-z-w) ≥ 0.95, and sintering aids (Fe2O3, MnO2, and CuO) were encoded into the input as a separate entity to control non-stoichiometric or defect-dominant ABO3 outputs as well as the microstructural distribution. During inference, the model weights remained frozen, and no additional training or fine-tuning was performed. The few-shot examples included in the prompt, therefore, act only as in-context demonstrations to guide the model’s reasoning, rather than serving as training data that updates the model parameters. The model was further prompted to yield a predefined multiphase combination containing R3m, Amm2, and P4mm phase symmetries in targeted fractions of 20%, 40%, and 40%, respectively, ensuring that the predicted outputs reflected realistic crystallographic phase competition rather than purely statistical token space interpolation. The phase fraction targets used above, as Strategy S7, were explicitly encoded as textual directives within the strategy string provided to the model, rather than being inferred from the few-shot chemistry d33 examples. Accordingly, the model leverages statistical heuristics derived from its training corpus to propose candidate compositions that are likely to satisfy the specified phase coexistence criteria. LLMs fundamentally operate on probabilistic token generation, and sometimes the predictions violate electroneutrality or produce oxygen-deficient perovskite variants that violate the tolerance limit of the host lattice. To overcome these issues, prompt engineering was employed to embed strict compositional and structural constraints, ensuring the LLM generated chemically valid outputs, as illustrated in Fig. 1b. Composition generation (in Claude 3.5 Sonnet) was carried out at a temperature set to “0” to maximize determinism (minimize stochasticity)23. However, minor variability may still arise due to inherent LLM inference limitations24. Importantly, repeated runs and physics-based filtering confirm that the model’s compositional ranking and selection of viable candidates remain robust despite such variability. Physics-based sanity checks include theoretical evaluations, such as charge balancing, A-site/B-site valence parity, and compliance with sintering-aid constraints. Prior to this, the curated compositions were categorized into two functional groups based on a performance threshold of d33 ~ 450 pC/N, enabling systematic segregation of high-confidence compositions from low-performing or chemically inconsistent predictions. Figure S4 shows the experimental difference between the electrical properties of both the compositions, each picked from the group. The details of the dual segregation analysis and their respective interpretations of each composition are discussed in Section S3 (S.I). To validate the reliability of the selected LLM (Claude 3.5 Sonnet), we performed a post-hoc comparative analysis against other frontier models (Figs. 1c and S6). Using the experimentally determined values of C1 and C6 (confirmed in Section S3) as ground truth, the analysis highlighted a critical divergence in reasoning capabilities. While DeepSeek-V3.2 and Gemini 2.5 Pro correctly identified general trends, they often relied on anchoring heuristics, simply penalizing low-performing compositions without identifying the root cause. In contrast, Claude 3.5 Sonnet demonstrated superior adherence to the complex system prompts, correctly prioritizing composition C1 based on the physical necessity of the Fe2O3 sintering aid encoded in the few-shot examples. The LLM-based approach presented here is intended to complement rather than replace traditional materials, informatics, or structured dataset machine learning methods. The main advantage lies in enabling meaningful exploration of data-sparse multi-doped composition spaces, such as modified KNN systems, where sufficiently large and well-curated composition-property datasets required for training conventional supervised ML models are currently unavailable. Material characterization and electrical properties Figure 2 demonstrates the phase evolution of overall composition as a function of temperature and content of Fe2O3. Figure 2a shows the selected XRD peaks at 2θ ~ (45°– 46.5°), highlighting the splitting of the perovskite peak for phase fraction. Pure orthorhombic phase (Amm2) is identified with I022/I200 ~ 1.7 while pure tetragonal phase (P4mm) corresponds to I002/I200 ~ 0.525. Figure S7 depicts the Rietveld refinement analysis of composition with 0.1 and 0.2 mol% Fe2O3. The approximate volume fractions of R3m:Amm2:P4mm calculated from Rietveld refinement are 5:50:45 and 18:48:34 for the compositions containing 0.1 mol% Fe2O3 (modified composition) and 0.2 mol% Fe2O3 (predicted composition), respectively. This is in alignment with the prompted strategy (Strategy S7) depicted in Table S1. The details behind the development of Strategy 7 is discussed in Section S1.Fig. 2Phase evolution and texturing behavior of KNN-BNKZT-SCZ ceramics as a function of Fe2O3 content.a Selected XRD peaks (2θ ~ 45–46.5°) showing phase coexistence and evolution among rhombohedral (R3m), orthorhombic (Amm2), and tetragonal (P4mm) phases. b XRD patterns of textured samples, indicating a maximum texture degree (>97%) at 0.1 mol% Fe2O3. c Temperature dependence of electromechanical coupling factors (k31 and kp). d–f Dielectric permittivity and phase transition behavior, demonstrating intrinsic multiphase coexistence and enhanced thermal stability at optimal Fe2O3 content. Optimal phase balance and microstructure achieved at 0.1 mol% Fe2O3 result in superior texture development and electromechanical performance. All textured compositions have 3 wt% NaNbO3 seeds. Phase evolution and texturing behavior of KNN-BNKZT-SCZ ceramics as a function of Fe2O3 content. a Selected XRD peaks (2θ ~ 45–46.5°) showing phase coexistence and evolution among rhombohedral (R3m), orthorhombic (Amm2), and tetragonal (P4mm) phases. b XRD patterns of textured samples, indicating a maximum texture degree (>97%) at 0.1 mol% Fe2O3. c Temperature dependence of electromechanical coupling factors (k31 and kp). d–f Dielectric permittivity and phase transition behavior, demonstrating intrinsic multiphase coexistence and enhanced thermal stability at optimal Fe2O3 content. Optimal phase balance and microstructure achieved at 0.1 mol% Fe2O3 result in superior texture development and electromechanical performance. All textured compositions have 3 wt% NaNbO3 seeds. Figure 2b shows the XRD pattern for its corresponding texture counterpart. The maximum texture degree (f %) is achieved with 0.1 mol% Fe2O3 (>97%), which is also evident from the low intensity of the (110) peak as a function of Fe2O3 content. Since Fe2O3 is an effective sintering aid for KNN compositions, its tendency to form an excess liquid phase controls the nucleation and growth kinetics as well as the surface energy for texture growth, achieving an optimal phase and microstructure around 0.1 mol% Fe2O326,27. The microstructure evaluation in Fig. S8 also validates this observation. In the absence of Fe2O3 (0% Fe2O3), the microstructure consists of relatively small textured grains, attributed to incomplete stage-III sintering and suppressed Ostwald ripening due to insufficient liquid-phase formation. In contrast, increasing the Fe2O3 content to 0.2 mol% results in the emergence of fine, randomly oriented grains between textured grains. This behavior is attributed to excessive volumetric nucleation within the random matrix, which competes with template-assisted planar nucleation. Also, excess liquid phase induced by higher Fe2O3 content promotes rapid homogeneous nucleation and partial dissolution of NaNbO3 templates, thereby suppressing templated grain growth and texture evolution28. This phenomena was profoundly observed in Fig. S11 (EDS analysis of cross-section textured surface). As the amount of Fe2O3 was increased to 0.2 mol%, templates started to dissolve in the matrix (Fig. S11c), consequently increasing the fraction of small random grains between large <001> textured grains (observed in Fig. S8). This reduces the texture degree (f ~ 93%), which leads to a concomitant decline in d33 and phase of impedance depicted in Fig. S12f after the texturing process. From the observation of the ratio of peak intensity in XRD, Rietveld fittings, and the broad ferroelectric-ferroelectric phase transition peak TROT ~ 50 °C in εr − T measurements (Fig. 2d), KNN-BNKZT-SCZ intrinsically possesses mixed phases. As the Fe2O3 content increased, it merged into a single peak in XRD, as well as a shift in the Curie peak (Tc) and TROT to higher temperatures are observed in Fig. 2d, e. The differences in the magnitude of εr are attributed from simultaneous effect of both phase (intrinsic) and microstructure (extrinsic) contributions. Since the ferroelectric-ferroelectric (TROT) phase transition in the case of random compositions is not significant, the contribution is most likely due to the change in the microstructural distribution, which can also be inferred from Fig. S8. With 0.1% Fe2O3 random composition, a more nonuniform and broader distribution of grains are observed compared to undoped and 0.2 mol% Fe2O3 doped composition, possibly leading to a higher number of possible states for polarization rotation and extension. Similar measurements were conducted on the textured counterpart depicted in Fig. 2f. Due to crystal anisotropy, the baseline of εr was reduced, showing a minimized effect of broad phase transitions, resulting in a more stable εr as a function of temperature (Fig. 2f). Moreover, the Curie peaks (after texturing) exhibited a more diffused transition with an increase in Fe2O3 concentration. This phenomena is most likely originating from chemical disorder introduced by excess doping4. Further, the absence of one-dimensional morphology of the NN templates in the cross-sectional EDS mapping (Fig. S11) is likely due to their dissolution during the sintering process, which may introduce chemical disorder and contribute to the observed diffuse behavior. Furthermore, to monitor the similar trends on the electrical properties of the textured samples, k31 and kp were measured as a function of temperature, as shown in Fig. 2c. Both coupling factors as well as the piezoelectric charge constant (d31, Fig. S13) indeed stay large up to 165 °C, near to Tc (170 °C), but show a modest decrease between 30 °C (R.T) and 60 °C. The fall in the magnitude is attributed to the deviation from the designed R-O-T ferroelectric-ferroelectric phase boundary (broad peak observed at R.T). Apart from the range, a consistent behavior was observed after 60 °C in Fig. 2c for kij, as well as in the magnitude of and d31 until Tc. Figure 3 shows the microstructural characterization of both the random and textured ceramics. The surface morphology of randomly oriented composition shown in Fig. 3a poses a nonuniform grain distribution along with a multicolor mosaic pattern in inverse pole figure (IPF) (Fig. 3c). However, with texturing, a more uniform and larger grain distribution was achieved, as shown in Fig. 3d. The grains are well-oriented along the <001> direction, as concluded from the electron beam backscattered diffraction (EBSD) map in Fig. 3e. The IPF map shown in Fig. 3f defines the crystal axis (<001> direction), which is well aligned with the sample normal (mapped along the “z” direction), further confirming a very high degree of texturing. However, a high texture degree (f ~ 97.3%) is confirmed through XRD, EBSD, and IPF mapping, which is attained due to both a high-quality NN seed crystal that is properly aligned and meticulously sintered with an optimized sintering profile.Fig. 3Microstructural analysis of random and textured compositions.Band contrast image of a random and d textured KNN-BNKZT-SCZ 0.1 mol% Fe2O3 taken along the “z” direction. EBSD spectra captured for b random and e textured KNN-BNKZT-SCZ 0.1 mol% Fe2O3 captured along the “z” direction with its corresponding scale. Inverse pole figure (IPF) measured for c random and f textured ceramics. Microstructural analysis of random and textured compositions. Band contrast image of a random and d textured KNN-BNKZT-SCZ 0.1 mol% Fe2O3 taken along the “z” direction. EBSD spectra captured for b random and e textured KNN-BNKZT-SCZ 0.1 mol% Fe2O3 captured along the “z” direction with its corresponding scale. Inverse pole figure (IPF) measured for c random and f textured ceramics. Figure 4a shows the impedance sweep of both random and textured compositions in the planar mode of operation, with D.C. poling at room temperature in a Si oil bath. Both the samples manifest a reasonable degree of poling, which can be established from the higher phase matching (θ ~ 70°). The bandwidth (fa − fr) of textured ceramics was calculated to be ~37.5 kHz, while that of random is ~35.5 kHz. The difference of 2 kHz can likely be attributed to the difference in grain size as well as the orientation. As the concentration of defects and un-passivated/dangling oxygen bonds increases at grain boundaries, it acts as a conductive pathway between electrodes, leading to reduced gap between open circuit (fa) and short-circuit (fr) conditions, which ultimately enhances the electromechanical coupling (kij)29. Furthermore, Fig. 4b, c demonstrates the polarization and strain measurements conducted at room temperature. The observed increase in remanent polarization (Pr) and strain (%) is likely associated with enhanced crystallographic anisotropy induced by the high degree of texture30. The significant difference in the increment in strain (%) compared to Pr after texturing are related to the mechanisms associated with mechanical and electrical displacement, respectively. With uniform alignment of domains (through texturing), macroscopic mechanical deformations (strain (%)) under external field maximizes due to the simultaneous effect of lower energy barriers for switching, positive contribution from non-180° domain switching, as well as lower pinning effect31. Further, uniform field distribution across grains reduces the effect of depolarizing fields and local field cancellations. However, the magnitude of Pr depends on the material’s ability to retain polarization after the removal of the external field. Meanwhile, the magnitude of the coercive field (Ec) associated with the composition remains unaffected by texture. However, the range of Ec lies approximately 6.2 kV/cm, which is comparatively lower than that of the undoped KNN composition (in K0.5Na0.5NbO3,~10 kV/mm) and can be described as “soft” piezo, resulting in easy domain switching and a higher magnitude of electrical properties2.Fig. 4Electromechanical and ferroelectric properties of random and textured KNN-BNKZT-SCZ ceramics.a Planar-mode impedance spectra after DC poling at room temperature, showing enhanced bandwidth and electromechanical coupling in textured ceramics. b, c Polarization–electric field (P–E) and strain–electric field (S–E) responses, revealing increased remanent polarization and strain due to crystallographic texturing, while the coercive field remains largely unchanged. Electromechanical and ferroelectric properties of random and textured KNN-BNKZT-SCZ ceramics. a Planar-mode impedance spectra after DC poling at room temperature, showing enhanced bandwidth and electromechanical coupling in textured ceramics. b, c Polarization–electric field (P–E) and strain–electric field (S–E) responses, revealing increased remanent polarization and strain due to crystallographic texturing, while the coercive field remains largely unchanged. The computationally (statistically) predicted electrical properties (specifically d33) and the measured intensities are influenced by several interconnected material parameters, which include the precise phase fractions, grain size distribution, and other microstructural features, the densification level, and the degree of chemical homogeneity within grains and throughout the bulk sample. Variations in any of these parameters from the predicted structure during synthesis and processing can significantly affect the piezoelectric response, leading to the discrepancy between the predicted (~660 pC/N) and experimentally measured values (440–500 pC/N). Characteristics of MME energy harvesting performance with developed ceramics To explore the ME performance of the composite with Ni and developed piezoelectric ceramics and to evaluate the difference in the performance between the random and its textured counterpart, the MME energy harvesting measurement was carried out at the second harmonic resonance frequency of the piezoelectric cantilever beam. The ME voltage coefficient was characterized using the standard method at 1 kHz off-resonance condition. Figure 5a depicts the direct ME voltage coefficient (αDME), which approximates the conversion of magnetic field to voltage. Considering the condition of perfect coupling between Ni and KNN, αDME can be estimated as32,331\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\alpha }_{{DME}}=\,\frac{\partial E}{\partial P}\,=\frac{2{.g}_{{ij}}.{\sigma }_{{ij}\,p}^{E}}{{H}_{{ac}}}$$\end{document}αDME=∂E∂P=2.gij.σijpEHacFig. 5Measurement of αDME of random and textured composition (0.1 mol% Fe2O3) and its comparison.a Experimental setup for the characterization of the direct ME coupling coefficient (αDME) of a cantilever-structured ceramic/Ni composite as a function of DC magnetic bias field. The photograph shows the sample mounted inside the Helmholtz coil for evaluating the MME harvesting characteristics. A cantilever-structured ME composite is clamped at one end, while a permanent magnet proof mass is attached to the free end. An AC magnetic field corresponding to the second resonance frequency is applied along the longitudinal direction of the sample, thereby exciting the second bending resonance mode of the cantilever. b Comparison of off-resonance αDME for the composites with both random and textured ceramics, and c comparison of αDME with reported Pb-free compositions37–40. Measurement of αDME of random and textured composition (0.1 mol% Fe2O3) and its comparison. a Experimental setup for the characterization of the direct ME coupling coefficient (αDME) of a cantilever-structured ceramic/Ni composite as a function of DC magnetic bias field. The photograph shows the sample mounted inside the Helmholtz coil for evaluating the MME harvesting characteristics. A cantilever-structured ME composite is clamped at one end, while a permanent magnet proof mass is attached to the free end. An AC magnetic field corresponding to the second resonance frequency is applied along the longitudinal direction of the sample, thereby exciting the second bending resonance mode of the cantilever. b Comparison of off-resonance αDME for the composites with both random and textured ceramics, and c comparison of αDME with reported Pb-free compositions37–40. In Eq. 1, gij and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma }_{{ij\; p}}^{E}$$\end{document}σijpE present the voltage coefficient and compressive stress tensor on the piezoelectric layer, and Hac is the applied AC magnetic field. Considering the ideal coupling condition between KNN and Ni plate, the induced stress can be of similar magnitude under a uniform magnetic field for both random and textured laminates. The resulting change in the total αDME can be correlated to the change in the gij of random and textured composition. The ratio of g33 (calculated as d33/ε0εr) between random and its textured counterpart (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{g}_{33}\left({Textured}\right)}{{g}_{33}\left({Random}\right)}$$\end{document}g33Texturedg33Random) in the thickness direction closely matches the corresponding enhancement (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{\alpha }_{{DME}}({Textured})}{{\alpha }_{{DME}}\,({Random})}$$\end{document}αDME(Textured)αDME(Random)) in a factor of ~2.25. Consequently, the enhancement of αDME in textured composition is directly governed by the change in the g33 (voltage coefficient), which also aligns well with Eq. 133. This indicates that g33, representing the voltage generation capability of the piezoelectric material, plays a critical role in determining the magnetic field to voltage conversion behavior Fig. 5c presents the comparison of the αDME reported for Pb-free magnetoelectric particulate composites and laminates. In general, under off-resonance conditions, cantilever-type ME composites exhibit relatively lower ME coupling characteristics than layered magnetostrictive–piezoelectric–magnetostrictive composites, which enable uniform transfer of magnetostrictive stress34. However, in the present study, despite the cantilever configuration in which stress transfer from the magnetostrictive material occurs only from one side, a higher level of αDME is achieved, demonstrating the superior piezoelectric properties of the material playing an important role. Further, laminate architecture is known to yield high voltage output (O/P) compared to particulate composites, which remain valid in the case of Pb-free systems also35. The 2–2 laminate approach proved to be more efficient, resulting in a higher magnetoelectric voltage coefficient (αDME) due to better strain transfer, and also eliminates the need of complex sintering process in case of particulate composites32,36. Meticulously designed independent phases (magnetostrictive and piezoelectric) bonded together are more favorable to prevent interdiffusion between constituent phases that can cause property degradation. The textured ceramic exhibits high αDME due to its high g33 compared to ME designed via Pb-free particulate composites, laminate structures from random compositions37–40. The characterization of MME energy harvesting performance was conducted at the second harmonic of the resonant frequency of the cantilever beam41. The vibration modes in cantilever structures, in the context of piezoelectric actuators, predominantly depend on the excitation frequency of the magnetic field. In the first harmonic bending vibration mode, the maximum displacement is concentrated at the end of the cantilever beam. In the second harmonics, the maximum vibrational displacement is observed at the center of the beam directed along the thickness direction21. Considering the second harmonics (189 Hz for random and 309.9 Hz for textured), the energy harvesting measurement was carried out under 10 Oe AC magnetic field, and a comparative plot of O/P power as a function of impedance is plotted in Fig. 6a. The maximum RMS O/P power for textured ceramics was calculated to be 57.3 μW at 182 kΩ. Correspondingly, the power for its random counterpart is 0.1 μW at 272 kΩ. RMS power was calculated from the RMS voltage at its corresponding load resistance. Total power density calculated from the physical dimension of the ceramic samples aggregates to 1 μW/cm3 in the case of random composition, while that of the textured counterpart is 705 μW/cm3. Figure 6b depicts the RMS voltage measured across multiple external load resistances. The O/P voltage exhibits a saturating trend and saturates at ~0.25 V for random composition and 4.9 V for its textured counterpart. The giant difference in the power density can be attributed from the difference in FoMs in random and textured ceramics. More specifically, the textured ceramics exhibit an enhanced d33 along with a reduced dielectric constant, which leads to an increased g33. As the FoM (d33 × g33) reflects both charge generation and voltage generation capabilities, this combination results in more efficient energy conversion and consequently improved output power in the piezoelectric energy harvester. As summarized earlier in Table 1, the energy-harvesting FoM is improved from 5.4 to 17.3 pm2/N by texturing, corresponding to an enhancement of 320%. As a result, the ME composites incorporating textured ceramics demonstrate an exceptionally enhanced MME energy-harvesting output performance. A comparison of the final O/P power density and voltage attained with multiple Pb-free ME systems is depicted in Fig. 6c42–44. The current textured composition epoxied with magnostrictive Ni foil resulted in very high-power density along with O/P voltage. However, their power density does not exhibit a significant difference in magnitude between the laminate composite and the independent piezoelectric layer depicted in Fig. 6b. The thickness of magnetotsrictive layer compared to the piezoelectric ceramic, single crystal, or ceramic/polymer composite is minimal, which helps with an efficient vibration transfer without affecting the final output power density. The acquired range of voltage and power density has a huge application in a broad spectrum of energy harvesting and wireless data transmission.Fig. 6Energy harvesting performance of MME structures prepared from random and textured compositions and their respective comparisons.a RMS power measured as a function of load resistance along with its respective power density (PD). Since the difference between the RMS power of random and textured compositions is too high to accommodate in a single graph, the axis are scaled accordingly. b RMS output voltage measured across multiple load resistances. c Comparison with reported Pb-free ME laminated devices based on its respective output voltage and power42–44.Table 1List of the electromechanical properties of random and textured ceramics as a function of sintering aid (Fe2O3) used, measured at room temperaturePropertiesRandomTextured0 mol% Fe2O30.1 mol% Fe2O30.2 mol% Fe2O30 mol% Fe2O30.1 mol% Fe2O30.2 mol% Fe2O3d33 (pC/N)350440500470620350εr310040003200227025002120k310.290.310.310.290.350.27g33 (10−3 Vm/N)12.712.417.623.42818.6d33 × g33 (10−12 m2/N)4.45.48.810.917.36.5ρ (%)999999999998f (%)−/−−/−−/−9797.393 Energy harvesting performance of MME structures prepared from random and textured compositions and their respective comparisons. a RMS power measured as a function of load resistance along with its respective power density (PD). Since the difference between the RMS power of random and textured compositions is too high to accommodate in a single graph, the axis are scaled accordingly. b RMS output voltage measured across multiple load resistances. c Comparison with reported Pb-free ME laminated devices based on its respective output voltage and power42–44. Table 1List of the electromechanical properties of random and textured ceramics as a function of sintering aid (Fe2O3) used, measured at room temperaturePropertiesRandomTextured0 mol% Fe2O30.1 mol% Fe2O30.2 mol% Fe2O30 mol% Fe2O30.1 mol% Fe2O30.2 mol% Fe2O3d33 (pC/N)350440500470620350εr310040003200227025002120k310.290.310.310.290.350.27g33 (10−3 Vm/N)12.712.417.623.42818.6d33 × g33 (10−12 m2/N)4.45.48.810.917.36.5ρ (%)999999999998f (%)−/−−/−−/−9797.393 List of the electromechanical properties of random and textured ceramics as a function of sintering aid (Fe2O3) used, measured at room temperature This work presents a paradigm shift in the discovery of application-specific piezoelectric compositions. By integrating AI-assisted extraction of complex dopant patterns with targeted refinement guided by human intuition. Other recent work in this area also points to the possibility of AI-guided experimental design and optimization, suggesting that a full materials discovery pipeline—including candidate proposal and process optimization—could be performed with a similar approach17. The pretraining corpora of commercial LLMs are not publicly disclosed. As a result, the extent to which domain-specific scientific literature, particularly in specialized areas, such as piezoelectric materials and microstructural characterization, is represented in their training data is unknown. This makes it difficult to rigorously assess the difference in generalizable knowledge internalization (effect of Bi3+ on covalency) and statistical interpolation (improper combination of additives affecting electroneutrality) and therefore constitutes an important open research question. This study establishes a scalable pathway for electromechanical performance, while validating functional relevance through device-level energy harvesting. In summary, KNN-based compositions with targeted electrical properties were identified using a collaborative human-AI (Claude Sonet 3.5) candidate generation strategy and refined through human-in-the-loop prior to synthesis. The selected ceramic exhibited a d33 ~ 440–500 pC/N in its random form, which was further enhanced to 620 pC/N through controlled texturing, exhibiting excellent thermal stability till 160 °C. To assess the MME performance, the output power of both random and textured ceramics were measured under second harmonic excitation. The ratio between the direct ME coefficient (αDME) of random and textured compositions were consistent with their corresponding g33 enhancement, while the textured composition delivered a maximum power density of ~705 μW/cm3, significantly higher than its random counterpart (~1 μW/cm3). Overall, the strong agreement between AI-guided predictions, experimental outcomes, and benchmarking against state-of-the-art Pb-free MME harvesters provides the promise of an effective pathway for application-specific material discovery for next-generation functional ceramics. Methods Candidate generation with LLM To accelerate the discovery of high-performance piezoelectric materials, we employed an LLM (Claude 3.5 Sonnet, Anthropic) to propose novel, chemically valid compositions guided by specific constraints16. The methodology involved few-shot learning principles, where the model was provided with 10 carefully selected reference composition-property pairs (compositions with their corresponding d33 values) to establish the relationships between chemical composition and piezoelectric performance. The reference compositions span d33 values from 231 to 680 pC/N, representing a diverse range of KNN-based piezoelectric materials with various dopants and secondary phases (detailed compositions provided in Supporting Information). This few-shot example approach leverages the model’s ability to generalize from limited examples and generate new compositions with potentially enhanced piezoelectric properties. Seven distinct generation strategies were systematically used to explore different avenues for composition optimization (Table S1). These strategies roughly fell into three categories: (i) direct elemental substitution (1–3), focused on isovalent or aliovalent substitutions; (ii) structural and relaxor modification (4–6), focused on local structural heterogeneity or relaxor behavior; (iii) targeted phase engineering (7), where the model was explicitly constrained to generate compositions yielding specific crystallographic phase fractions. Each strategy was implemented through domain-expert-designed prompts to guide the generation process. The complete prompt structure included the few-shot examples, specifying the strategy, and output formatting requirements45. To ensure the generated compositions remained physically realistic and synthetically feasible, the model was provided with detailed system prompts that established strict compositional templates and constraints. This system prompt was developed through iterative prompt engineering, following several trial-and-error exercises, to validate that the LLM could generate chemically valid compositions that adhered to proper stoichiometric rules and structural requirements. These constraints ensured adherence to stoichiometric requirements (1-x-y-z-w) ≥ 0.95, electroneutrality, and limitation of oxide dopants to Fe2O3, MnO2, and CuO, and preservation of perovskite structure integrity (ABO3 framework). This systematic approach generated 147 unique compositions (7 strategies × 21 compositions each), creating a diverse library of candidates. Each generated composition included a complete chemical formula following given prompts, an estimated d33 value based on compositional analysis, and a chain of thought for the predicted enhancement in property, incorporating established structure-property relationships. Detailed prompting is illustrated in Section S1, and the respective strategies employed during prompting are outlined in Table S1. Material synthesis and characterization The most promising candidate composition identified through the LLM model, 0.96(K0.5Na0.5)(Nb0.94Sb0.06)O3-(0.03)(Bi0.5Na0.4K0.1)(Zr0.96Ti0.04)O3-0.01(Sr0.7Ca0.3)ZrO3 – x %Fe2O3 (x = 0, 0.1, and 0.2) (KNN-BNKZT-SCZ) was synthesized via a conventional solid-state synthesis route. High-purity precursor materials Nb2O5 (Thermo Scientific 99.9%), K2CO3(VWR, 99%), Na2CO3 (Sigma Alderich, 99.5%), Sb2O5 (Alfa Aesar, 99.6%), Bi2O3 (Thermo Scientific, 99%), ZrO2 (Alfa Aesar, 99.7%), TiO2 (Alfa Aesar, rutile, 99.8%), Sr2O3 (Sigma Alderich >99.9%), and Ca2CO3 (Sigma Alderich >99%) were weighed and ball-milled for uniform mixing in ethanol. Further, the powders were dried and calcined at 850 °C for 6 h in a muffle furnace. The calcined powders were ball-milled with x mol% Fe2O3 to achieve the optimum grain size and uniformity for grain growth. Ball-milled powders were dried, pelletized, and sintered with a multistep sintering profile to achieve ~99% theoretical density (T.D). The temperature for sintering was raised to 1190 °C at a rate of 5 °C/min and then cooled rapidly to 1180 °C at a rate of 10 °C/min, where it was maintained for 10 h30,46,47. Density measurements were carried out using Archimedes' method. Textured ceramics were prepared by tape casting the slurry using a doctor blade at a speed of 4 cm/s. The slurry was prepared using randomly oriented calcined powders mixed with NN templates (3 wt% relative to the random powder), along with binders and ethanol. The amount of template was considered based on reported literature and prior optimization studies on multielement doped KNN compositions, which indicate an optimal range of 3–4 wt% of total random powder taken47–50. Prior to the tape casting, the NN seeds were prepared via molten salt synthesis followed by topochemical microcrystal conversion51. The average aspect ratio of NN seed observed through FE-SEM measured up to 15 μm × 15 μm with a thickness of a few 100 nm. Multiple layers of tape cast sheets were further laminated and sintered using a similar temperature profile to that used for random ceramics. All the ceramics were machined according to the IEEE standard and painted with Ag paste, then cured for measurement52. The phase of the ceramics was characterized through X-ray diffraction (XRD, Malvern Panalytical, Empyrean III). Texture degree was calculated through Lotgering formula as47:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L.F(f)=\left(P-{P}_{0}\right)/\left(1-{P}_{0}\right)$$\end{document}L.F(f)=P−P0/1−P0where P is the ratio of the peak sum of peaks in the preferred orientation of textured ceramics, while P0 is a similar parameter for random ceramics. Microstructural characterization and mapping of grains for preferred orientation were done with FE-SEM and Electron beam backscattered diffraction (EBSD) with Thermo Fischer Scientific, Apreo, using the software platform of Oxford Instruments. Elemental distribution was examined by energy-dispersive X-ray spectroscopy (EDS) using the same instrument. Crystal structures were visualized and analysed with CrystalMaker 11, and all graphical data were processed and plotted using Origin 2025. Characterization of electromechanical properties Impedance spectra were observed with an Impedance Analyzer (Keysight E4990A), which was further used to calculate the electromechanical coupling factor via resonance- antiresonance positions. Relative permittivity as a function of temperature (εr − T) was measured through an LCR meter (Keysight E4980A) synced with a temperature-controlled chamber. Thermal stability measurements were performed using the same setup. The piezoelectric charge constant (d33) was measured using a Berlincourt d33 meter, and the voltage coefficient (g33) was calculated from the ratio of d33 to the relative permittivity (εr), i.e., g33 = d33/(εrεo) Prior to the measurements, all samples were DC-poled in silicone oil at room temperature under an electric field of 4 kV/mm. Characterization of MME energy harvesting performance To evaluate the improvement in MME harvesting device performance of the fabricated piezoelectric ceramics, ME composites were prepared by bonding a magnetostrictive Ni plate to the piezoelectric samples (schematics shown in Fig. S3b). The Ni plate (nickel foil, 99.5% purity, annealed, 0.25 mm thickness, Alfa Aesar, USA) was machined to dimensions of 55 mm × 6 mm × 0.25 mm. Two types of piezoelectric ceramics were used: random-oriented and [001] textured samples with approximately identical dimensions of 4 mm (width) × 22.5 mm (length) × 1 mm (thickness). Each ceramic sample was bonded onto the Ni plate using DP460 epoxy (3 M, MN) adhesive. During assembly, the ceramic was attached at a position 1 cm away from the clamped end of the Ni plate. Proof mass was attached to the end of the cantilever to provide inertia during the measurement. The distance between the proof mass and the free end of the cantilever was approximately 6 mm, with slight variations depending on sample length, for both the randomly oriented and textured samples. The inward position of the proof mass from the cantilever end was decided based on the maximum output voltage at the 2nd resonance. The piezoelectric ceramic was bonded directly onto the Ni plate, which served as the cantilever. The structure was directly clamped at one end of the Ni plate without the use of any additional substrate. The ME composite, consisting of a ceramic bonded to a single Ni plate, was placed at the center of the electromagnet and Helmholtz coil with its in-plane direction aligned parallel to the magnetic field. A DC magnetic field was supplied through the electromagnet using a DC power supply, while an AC magnetic field was generated by driving the Helmholtz coil with a lock-in amplifier (SR865A, Stanford Research Systems, USA) combined with a high-speed bipolar amplifier. Schematics of the measurement setup is represented in Fig. S3a. The superimposed magnetic fields were applied along the transverse (in-plane) direction of the ME composite. The induced voltage from the ME composite was amplified using a differential amplifier and subsequently detected by the lock-in amplifier. The direct ME coefficient (αDME) was obtained under an off-resonance AC magnetic excitation of 1 kHz as a function of DC bias magnetic field. Supplementary information Supplementary Information Transparent Peer Review File Supplementary Information Transparent Peer Review File