Data-driven concavity engineering for magnetic activation in rigid microspheres Authors: Chang Zhang, Longjun Rao, Lishan Wu, Ruixuan Zhang, Jiazhuan Qin, Meichen Wen, Ke Pei, Hanwen Cheng, Chongyun Liang, Wenbin You, Renchao Che Published: 2026 ABSTRACT AbstractConcave architectures harness geometric curvature to amplify interfacial activity and local fields, presenting encouraging prospects for catalytic, optical, and electromagnetic applications. However, their controllable synthesis in rigid metals remains challenging due to pronounced anisotropic surface energy and crystallographic constraints. Departing from conventional thermodynamically driven mechanisms, we propose a broadly applicable stress manipulation strategy during droplet evaporation to induce and regulate surface concavity in rigid microspheres. Such deliberately engineered concave configuration, guided by machine learning optimization, enables enhanced magnetic activity in electromagnetic systems. Reprogrammed magnetic domains orchestrate localized energy redistribution and magnetic activation, meanwhile effectively reducing spatial occupancy. The optimal FeCo alloys outperformed commercial magnetic powders, demonstrating a 28% increase in high-frequency magnetic response and a 79.5% reduction in density, while maintaining high saturation magnetization (236.6 emu/g). This mechanism successfully decouples magnetism from density, allowing for better design of lightweight, multifunctional materials.Subject terms: Magnetic properties and materials, Structural properties Introduction Precisely regulating active sites plays a crucial role in tailoring reactivity and selectivity of functional materials across diverse applications, from catalysis to electronics1–8. Yet the intricate structural requirements pose persistent challenges in both design and synthesis. Recently, extensive efforts have been devoted to manipulating reaction kinetics and surface energy to expose active sites9,10. For example, polyhedral nanostructures with high-index facets can accelerate catalytic reactions but often suffer from synthetic complexity and structural instability11–13. Another strategy involves synthesizing thermodynamically stable epitaxial morphologies to enhance anisotropy and interfacial activity14–16, though increased spatial occupancy inevitably results in higher material density. Hollow and porous structures have therefore been widely applied to reduce material density, but often at the expense of losing active sites17,18. This structural-functional trade-off is particularly evident in magnetic materials, where disrupted magnetic domain continuity in porous frameworks leads to appreciably attenuated magnetic responses19,20. Thus, reconciling active-site exposure with structural lightweighting remains critical for advancing next-generation multifunctional materials. Curvature manipulation offers an effective pathway to decouple magnetism from density. Concave architectures, in particular, can redistribute local strain and energy to enhance magnetic activity and simultaneously reduce spatial occupancy21–26. However, fabricating concave structures in rigid metallic systems remains challenging due to limited atomic mobility and strong surface energy anisotropy27. Overcoming these kinetic and thermodynamic constraints typically relies on sacrificial templating or harsh acid etching, which limits scalability and environmental sustainability28–30. Crucially, deliberately engineered concave architectures should be applicable across diverse material systems and remain effective even when combined with porous and hollow features. Nevertheless, such integration often introduces greater structural complexity. Fortunately, machine learning (ML) provides a powerful solution for navigating these high-dimensional design spaces, enabling efficient structural screening and accelerated optimization in complex systems31–37. Given these, achieving universal concavity synthesis in rigid structures, coupled with ML-accelerated design, are both significant and challenging. To reconcile active-site exposure with structural lightweighting, we propose a stress-guided synthesis strategy that enables precise modulation of surface concavities and pores in rigid microspheres for electromagnetic applications. Complex structures with varying concavities and porosities were efficiently designed using ML, and their magnetic domain activity and high-frequency magnetic responses were evaluated through micromagnetic simulations. The optimized structure featured up to 79.7% active magnetic regions compared with the original counterpart. By integrating data-driven design with concave engineering, the resulting FeCo microspheres exhibited superior magnetic properties, accompanied by a notably low density (0.92 g/cm3) compared to commercial counterparts (4.45 g/cm3), effectively overcoming the long-standing trade-off between magnetism and density. Furthermore, this synthesis strategy was experimentally validated across multiple metal alloy and oxide systems, underscoring its broad applicability. Taken together, this work provides a viable and scalable pathway for the design and synthesis of concave structures in rigid systems, while inspiring valuable insights for applications in catalysis, drug delivery, and superhydrophobic coatings. Results and discussion Theoretical design and optimization of concave structures Smart wearables in close contact with human skin have raised concerns regarding localized electromagnetic interference (EMI)38. Functional materials designed to mitigate EMI are required to exhibit high absorption efficiency and lightweight characteristics for practical applications39. Consequently, integrating miniaturized circuits into thin patches enables wearable devices with excellent conformability and effective electromagnetic wave absorption (EMWA), ensuring the stability and safety of electronic systems (Fig. 1a). A key determinant of EMWA performance lies in the internal structure of magnetic material. As shown in Fig. 1b, transitioning from bulk magnets to powders improved flexibility and reduced weight. However, further size reduction often compromises active magnetic regions, leading to weak magnetism or even superparamagnetism. Thus, structural redesign is essential to prevent deterioration of magnetic activity and ensure stable performance. Although polyhedral or spiky structures expose more active sites, they contribute little to reducing structural mass and may even increase material density owing to the accumulation of high-curvature features. To overcome this dilemma, we propose a synergistic design that combines surface concavities and pores to maximize active magnetic regions, meanwhile reducing material density. Motivation and implementation routes for active magnetic region engineering in concave microsphere. a The extensive application of wearable devices is driving increasing demand for advanced electromagnetic (EM) control technologies. Aiming for lightweight, thin, and multi-frequency EM control, a generation of flexible functional magnetic patches is being developed. b Traditional EM functional materials exhibit strong magnetism but are limited by high density, active magnetic regions engineering enables both lightweight design and efficient EM control. c The screening process of active regions: identifying inert magnetic domains under high-frequency excitation by micromagnetic simulations and iteratively refining the magnetic configuration, the proportion of active magnetic domains can be significantly improved. d For lightweight design, increasing concavity decreases the effective density at a given porosity. For a thin thickness design, higher relative permeability reduces the optimal matching thickness at different frequencies. For flexible design, the optimized magnetic film sustains higher stress over a broader strain range than a conventional magnetic film. The curves are schematic and intended to illustrate the expected design trends. Density functional theory (DFT) calculations have been widely employed to identify active sites in functional materials, particularly in catalytic systems, by evaluating adsorption energies and reaction energy barriers40,41. Similarly, micromagnetic simulations (MS) were utilized to investigate magnetic domain dynamics and permeability characteristics under high-frequency magnetic fields (Supplementary Fig. 1). Herein, we first simulated the evolution of magnetic domains in spherical, convex, and concave morphologies. The results showed that both convex and concave geometries enhance magnetic domain activity compared to regular spheres, thereby achieving higher permeability (Supplementary Fig. 2). This behavior can be attributed to the disruption of the uniform demagnetizing field in convex and concave morphologies, causing a spatial redistribution of magnetic free energy—primarily magnetostatic and exchange energies. The resulting reduction in local magnetic stiffness allows magnetic moments to rotate more easily under an external field, thereby promoting domain evolution. As a result, convex and concave structures exhibit more active magnetic-domain dynamics. We then quantified active magnetic domains by tracking dynamic moment rotation over a single period. Active magnetic domains (Δθ > 1°) were identified by evaluating the contribution of magnetic moment rotation angles to the total magnetic energy (Supplementary Fig. 3)42. Incorporating concavities and pores removed inert regions and elevated the active magnetic domain fraction from 52.9% (initial structure) to 79.7% (Fig. 1c). Notably, the optimization process was guided by active region analysis, ensuring the preservation of overall domain count and enhancing permeability. Theoretically, these features offer distinct structural advantages, including reduced overall density and improved permeability at low thicknesses. Moreover, surface concavities also enhanced interfacial anchoring with the matrix, thereby optimizing the tensile properties (Fig. 1d and Supplementary Fig. 4). Maximizing the number of active sites involves tuning multiple variables, including porosity, surface concavity, and composition, making exhaustive exploration via trial-and-error simulations or experiments impractical. Therefore, we leveraged ML models trained on MS-derived data (630 sets) to accelerate structural design of magnetic spheres, with a particular focus on surface concavity (Fig. 2a)42. Five-fold cross-validation was used to evaluate the generalization performance of the various models (Supplementary Figs. 5, 6 and Tables 2, 3). Among various tested ML algorithms, the XGBoost model yielded strong predictive capability for estimating the peak imaginary permeability (μ′′). To be specific, the average R² values for training (0.837 ± 0.012) and test (0.779 ± 0.055) sets were in close agreement. Notably, test-set performance showed limited variation across folds, indicating that the model maintains stable predictive behavior and generalizes reliably to unseen data (Supplementary Fig. 7). These results all demonstrate the stable generalization of XGBoost model in capturing structure–magnetic property relationships (Fig. 2b). Shapley additive explanations (SHAP) analysis further revealed that the surface concavity as the most influential structural feature (Supplementary Fig. 8), while Pearson correlation analysis confirmed strong coupling between porosity, concavity, and permeability (Supplementary Fig. 9). In addition, we constructed a Fe7Co3 alloy database to examine the model’s predictive accuracy by restricting the descriptors to porosity and concavity42. Five-fold cross-validation shows stable performance (R² ≈ 0.64-0.75), but the overall accuracy decreased compared with the full-descriptor model (Supplementary Fig. 10). This decline reflects the intrinsic influence of composition on magnetic behavior, indicating that compositional descriptors are necessary for achieving higher predictive accuracy. Active magnetic region optimization route via micromagnetic simulation–machine learning integration. a Workflow for guiding magnetic structure optimization. Key design parameters including elemental composition, surface concavities and porosity. A library of 630 alloy configurations and permeability were generated by combining 15 surface functionalized structures with various concave ratios, 7 porosity, and 6 elemental compositions. Multiple machine learning algorithms were used for modeling and analysis. Based on the training effect of the model, 9 structures were selected, and the optimal magnetic structure was identified through domain knowledge. b Correlation map from the XGBoost model. c The predicted peak permeability of Fe7Co3 alloy. d–f Radar plot of μ′ ratios for spherical structures with different porosity and concavities relative to the initial Fe7Co3 sphere based on micromagnetic simulation. In the 2–8 GHz range, values > 1 indicate enhanced permeability, while values < 1 indicate reduced permeability (d), porosity (e), concavity (f), and concavity&porosity. The optimal structures predicted by the machine-learning model differ not only in concavity and porosity but also in elemental stoichiometry. In particular, an Fe/Co ratio of 7:3 was identified as the optimal magnetic response. We therefore further evaluated the μ” of the optimal structures against that of the original spheres in the 2–8 GHz range at this composition. These results highlight that, beyond empirical optimization, the data-driven framework enables exploration of a broader geometry–composition space and supports a more predictive route for materials design (Fig. 2c and Supplementary Figs. 11–13). In Fig. 2d, the effect of porosity on permeability is non-monotonic and reflects a balance between two competing effects: at low porosity, pores enhance local demagnetizing-field inhomogeneity and shape anisotropy, thereby promoting magnetization rotation, whereas excessive porosity reduces the effective magnetic volume fraction and weakens exchange coupling, leading to a suppressed overall magnetic response. In addition, surface concavities exhibited a stronger enhancing effect, peaking at ~ 40% concavity coverage. Further increasing the concavity, however, damaged active sites and reduced magnetism (Fig. 2e). Those results reveal that the magnetic response is governed by the coupled effects of porosity and surface concavity. In Fig. 2f, the highest magnetic response was predicted within a structural window defined by a concavity ratio of 40–50%, a porosity of 5–10%. Consequently, the optimal structure was the result of data-driven prediction under coupled geometric–compositional optimization. This result further demonstrates that the machine-learning model can effectively capture the geometry–magnetism relationship and identify high-performing, non-intuitive structures. Notably, the forward model that links geometric descriptors to magnetic properties provides a foundation for developing an inverse design framework. In future work, integrating generative models and optimization databases could enable mapping desired magnetic responses back to optimal structural configurations, thereby extending the potential of this approach for discovering magnetic materials. Controllable synthesis of concave microspheres To implement this structural optimization, we introduced surface concavities at the precursor stage to overcome crystallographic constraints in magnetic microspheres. Using competitive evaporation of a binary solvent system (water and ethylene glycol), we established non-equilibrium interfacial stresses during droplet evaporation, which drove controlled surface deformation and induced concavities in rigid microspheres. Furthermore, we employed a simulation approach (ANSYS Fluent) to elucidate the mechanism of concavity formation during the droplet evaporation process. The results indicate that anisotropic evaporation behavior and internal stress redistribution collectively drive concavity formation, offering theoretical guidance for morphological control43. As shown in Fig. 3a and Supplementary Movie 1, droplet volume progressively decreased, accompanied by the emergence of distinct surface concavities. During the initial stage, rapid water evaporation induced outward ion migration, forming a semi-solidified shell. Meanwhile, the slowly evaporating ethylene glycol remained trapped inside the droplet, increasing internal pressure (Supplementary Fig. 14). Once the pressure differential reached a critical threshold, localized shell collapse released the built-up stress, forming surface concavities. This process was kinetically frozen rather than thermodynamically equilibrated, with the structure trapped in a non-equilibrium state governed by dynamic stress evolution. Velocity vector field analysis revealed pronounced inward contraction at the droplet periphery, meanwhile pressure field mapping confirmed considerable internal stress accumulation, consistent with our predictions (Fig. 3b). Varying ethylene glycol concentration effectively modulated surface concavity, demonstrating the tunability of this evaporation strategy (Fig. 3c and Supplementary Fig. 15). Thermodynamically, concavity formation increased surface energy but simultaneously reduced elastic strain energy, resulting in a net decrease in total Gibbs free energy. In other words, when the internal stress exceeded a critical threshold, localized concavity formation became a thermodynamically favorable pathway for energy minimization (Fig. 3d). Moreover, by comparing the properties of ethylene glycol with other solvents, we further highlight its unique advantage in modulating stress during evaporation (Supplementary Table 4). These findings demonstrate that ethylene glycol plays a crucial role in regulating the evaporation behavior and governing the formation of surface concavities. Controllable synthesis pathway of concave and porous alloy microspheres. a Volume evolution process of droplets during high-temperature evaporation by Ansys Fluent simulation. b Velocity vector and pressure distribution cloud maps during evaporation. c Correlation between ethylene glycol content and surface concavities. d Schematic of the aerosol-assisted drying process: metal oxide form via droplets evaporation, followed by carbon removal and deoxygenation under high-temperature treatment, ultimately yielding concave and porous alloy microspheres. e SEM images of magnetic alloys with varying surface concavities. f Molecular dynamics simulation based on Fe5Co4Ni, revealing atomic diffusion behavior. g, h Diffusion rates of Fe, Co, and Ni atoms as a function of temperature, derived from the Arrhenius equation. Our results demonstrate that controlling stress distribution during droplet evaporation and optimizing the reduction process, the surface concavities and porosity of microspheres can be effectively tuned. Computational simulations provided insight into stress evolution during droplet evaporation, guiding the successful synthesis of rigid microspheres with varying degrees of surface concavity (Fig. 3e and Supplementary Fig. 16). Specifically, the surface coverage of concavities across microspheres could be programmed by adjusting solvent composition and evaporation dynamics. A detailed analysis confirming the formation of targeted alloys is provided in the Supplementary Information (Supplementary Figs. 17–19). To probe atomic-level mechanisms governing structural evolution during thermal reduction, molecular dynamics simulations were conducted to examine the diffusion behavior of Fe, Co, and Ni atoms (5:4:1) across a range of temperatures (Fig. 3f). All elements exhibited thermally activated diffusion, with Fe displaying slightly higher mean square displacement and diffusion coefficient than Co and Ni (Fig. 3g, h and Supplementary Fig. 20). The similar diffusion behavior of Co and Ni suggests their cooperative involvement in alloy framework formation, whereas the relatively faster migration of Fe may drive atomic redistribution. This behavior is likely related to the higher atomic fraction and comparatively weaker interatomic bonding of Fe44. Besides, the synthesis route is scalable through a continuous spray-drying process using low-cost precursors, enabling large-scale alloy powder production. These results reveal intrinsic diffusion heterogeneity among constituent elements and provide theoretical insights into thermally induced reorganization processes in magnetic alloys. The universality of concave engineering was also validated. As shown in Fig. 4a, we extended the synthesis to a broad range of metal systems, including main-group aluminum (Al) and transition metals such as manganese (Mn) and copper (Cu). Experimental results demonstrated the applicability of this strategy to construct both pure metals and high-entropy alloys (Fig. 4b). Scanning electron microscopy (SEM) and energy dispersive spectroscopy (EDS) mapping confirmed the formation of concave features across diverse metallic compositions (Fig. 4c and Supplementary Fig. 21). The universality validation of concave engineering. a A library with diverse concave microspheres (combination of vary metals, including transition metals, main group metals and precious metals). b Schematic atomic models of different types of concave microspheres, including metal elements, metal oxides, multicomponent alloys, and high-entropy materials. c–j EDS mapping of concave microspheres (c) Cu (d) Al2O3 (e)Fe3O4 (f) CoNi (g) FeCo (h) FeCoNi (i) FeCoNiMn (j) FeCoNiMnCu. scale bar, 1 μm. The corresponding elemental composition ratios also validated the compositional accuracy of alloy microspheres and universality of the synthesis method (Supplementary Table 5). Notably, the formation of the concave structure originated from the precursor oxide stage, indicating suitability for metal oxide synthesis—as demonstrated by the successful fabrication of concave Al2O3 and Fe2O3 microspheres. We further incorporated other transition metals into the FeCoNi matrix to construct quaternary and quinary alloys, which were similarly synthesized with morphologically controllable concave microspheres. X-ray diffraction (XRD) analysis confirmed their single-phase crystalline structures (Supplementary Fig. 22). These results collectively highlight the generality and versatility of stress-regulation strategy for the morphological control during droplet evaporation. Notably, concavity and porosity originate from distinct stages of structure evolution. Concavity is formed during solvent evaporation, whereas porosity mainly develops during oxide-to-alloy reduction through volume shrinkage and diffusion-induced void formation. This difference allows porosity to be tuned, to some extent, independently of concavity. A detailed pore characterization is provided in the Supporting Information (Supplementary Figs. 23–25). The synergistic integration of pores and surface concavities effectively reduced overall density. In summary, guided by theoretical simulations, we developed a universal, designable synthesis strategy for concave magnetic microspheres, establishing a solid foundation for subsequent studies on their properties. Verification of magnetic activity in concave alloys To elucidate the impact of active regions, we systematically investigated the magnetic behavior from both microscopic and macroscopic perspectives. Firstly, considering that energy is closely related to activity, magnetic energy mapping provides critical insights into the distribution of active magnetic regions45. The results revealed that magnetic energy predominantly accumulated in regions with high surface curvature (Fig. 5a). For regular spheres, the energy was concentrated at the poles. With the introduction of pores, energy was primarily distributed near pore boundaries. Further incorporation of surface concavities intensified energy accumulation around concave edges, indicating that local geometric heterogeneity strongly influenced magnetic energy distribution. Time-resolved analysis (0-200 ps) showed a universal dynamic pattern across all structures: an initial rapid energy release, followed by progressive absorption and amplification (Supplementary Fig. 26). Notably, the total magnetic energy of the concave-porous structure nearly doubled that of its uniform counterpart, indicating an excitation or metastable state. These localized features introduced perturbations in magnetic domains, thereby enhancing magnetic response and energy dissipation efficiency. Effect of active magnetic regions on magnetic properties. a Magnetic domain energy distribution of different magnetic sphere structures. b Optimization process of magnetic domains and corresponding periodic evolution. As the proportion of active magnetic domains increases, domain dynamics within each cycle intensify, leading to enhanced magnetic loss ability. c Magnetic interactions in different microstructural regions (smooth arcs, pores and concavities) of FeCo alloy via electron holography, revealing the distribution and interaction of stray fields across each region. d Using region C1 as a reference, the relative magnetic field intensities of C2, C3, and C4 were quantified. e Ratio of tan μ between optimized FeCo microspheres and dense FeCo spheres in 2-8 GHz. f Comparison of permeability (μ) between the optimized magnetic alloys and other magnetic materials. g Comparison of density and magnetic loss (tan μ) between the optimized magnetic alloys and conventional counterparts. Moreover, we explored the evolution of active domains and vortex dynamics under high-frequency magnetic fields (Fig. 5b). Concavities and pores increased the fraction of active magnetic domains and intensified vortex-related behaviors such as formation, migration, coalescence, and annihilation. These processes promoted local spin exchange and damping, contributing to efficient magnetic energy absorption and dissipation (Supplementary Fig. 27). To be specific, domain walls cannot follow the rapidly oscillating external field in the GHz regime, and the magnetization process is therefore entirely governed by magnetic-moment rotation. The resulting magnetic loss arises predominantly from natural ferromagnetic-resonance dissipation associated with magnetization precession, as well as relaxation losses induced by local magnetic anisotropy. These results reveal that the synergistic design of concavities and pores enhances magnetic performance by reshaping energy redistribution, elevating excitation states, restructuring domain configurations, and regulating vortex dynamics46. Apart from simulations, microstructural analysis of active regions was conducted experimentally. Off-axis electron holography was used to characterize local magnetic interactions (Fig. 5c). Specific regions—including inter-shell boundaries, pore edges, and concave sites—were analyzed. Localized magnetic field distributions were visualized by reconstructing phase maps, which clearly revealed coupled magnetic induction lines in all analyzed regions. The corresponding electron phase shifts were then used to calculate local magnetic field strength based on the following equation: where e is the electron charge, ℏ is the Plank constant (6.62 × 10−34 J/S), T is the sample thickness, L is the path length of the reference line, and B is the magnetic field strength. With all other parameters constant, variations in phase reflection correlated with the local magnetic flux. As shown in Fig. 5d and Supplementary Fig. 28, local field strength in concave regions was significantly higher than in reference areas, with a moderate enhancement evident near pore boundaries. This phenomenon can be attributed to curvature-induced flux convergence, which concentrated magnetic induction and increased energy density. These findings were consistent with the MS results. Moreover, magnetic flux lines penetrated into pore interiors, while localized concavities generated multiple magnetically active regions at the microscale, promoting rapid domain reconfiguration and improving magnetic response capability47. Although the pores may reduce the effective magnetic volume, the concave architecture introduces strong local demagnetization fields and high-curvature regions that dominate the magnetic response. As a result, the concave morphology effectively compensates for the adverse effects of porosity, enabling enhanced magnetic loss behavior. The macroscopic magnetic properties were also examined to verify enhanced magnetic activity. Vibrating sample magnetometer (VSM) measurements showed that optimal FeCo alloys maintained a desirable Ms (236.6 emu/g) compared with dense counterparts. This change is primarily attributed to enhanced magnetic anisotropy induced by local curvature. Although a slight decrease in Ms resulted from disrupted magnetic pathways, the elevated Hc improved EMI resistance and expanded operational frequency windows (Supplementary Fig. 29). Complex permeability measurements further confirmed this enhanced magnetic response capability. The optimized alloys displayed significantly higher real (μ″) and imaginary (μ″) permeability values, along with an increased magnetic loss tangent (tan δ) across 2-8 GHz (Fig. 5e). These values were competitive with, or superior to those of conventional magnetic alloys and composites (Fig. 5f, Supplementary Fig. 30a and Table 6). The measured μ″ curve exhibit similar trend and peak positions with micromagnetic simulations, further confirming the reliability of the simulation results (Supplementary Fig. 31). Similarly, the concave-engineered FeCo alloys achieved a 10%-28% enhancement in magnetic loss (tan μ) within the 1–4 GHz band compared with commercial carbonyl iron powders, demonstrating strong potential for 5 G communication and wireless networking applications (Supplementary Fig. 32). Remarkably, this improved performance was achieved while maintaining strong magnetic properties at a low bulk density (Fig. 5g and Supplementary Fig. 30b), underscoring a favorable balance between lightweight design and high-frequency magnetic loss capability. To further validate the mechanism that surface concavity enhances magnetic activity, we quantified the concavity parameter, including coverage, depth, roughness, and curvature, through multiple measurements. Statistical analysis of SEM quantitatively resolved microspheres with different concavity coverage (Supplementary Figs. 33–36). A clear rise-then-fall relationship of concavity coverage on μ″ is observed: μ″ increases slowly at low coverage (20–25%), reaches a pronounced maximum at 45–50%, and subsequently decreases as coverage continues to increase. Both the correlation trend and the peak position are in close agreement with the simulation predictions (40–50%). This behavior reveals that high concavity coverage promotes magnetic loss, while excessive concavities interrupt the continuity of effective magnetic domains and weaken the overall response (Supplementary Fig. 37). In addition, atomic force microscopy (AFM) was employed to quantify the concavity depth and roughness (Supplementary Figs. 38–40). Low-coverage samples contain relatively deep but sparsely distributed concavities with high roughness, whereas high-coverage samples exhibit more uniform depths and reduced roughness. Notably, 3D tomography provides microscopic, spatially resolved curvature information of microspheres, and BET surface area analysis reveals the overall concavity distribution over large statistical regions (detailed analysis is provided in the Supporting Information, Supplementary Figs. 41, 42). Together with the measured permeability results, these findings consistently demonstrate that concavity features are not only well-defined and quantitatively measurable but also exhibit a strong and physically meaningful correlation with magnetic enhancement. Therefore, concavity engineering offers a practical and effective strategy for boosting magnetic activity. The electromagnetic performance and application potential The active magnetic region design strategy enabled data-driven structural design and synthesis, enhanced magnetic response while maintaining low material density. These properties were evaluated under practical electromagnetic (EM) scenarios. Reflection loss (RL) curves, representing the ability to absorb EM waves from electronic devices, were first calculated across the 2–8 GHz band. The optimal FeCo alloy achieved a maximum absorption efficiency of 99.1% at 4.7 GHz, compared with 92.1% for its dense counterpart at the same frequency (Fig. 6a and Supplementary Fig. 43). At a thickness of 2.2 mm, it maintained over 60% absorption across a 4.2 GHz effective absorption bandwidth (EAB), whereas dense alloys failed to meet this threshold due to poor impedance matching (Fig. 6b). These results indicate that concave engineered alloys possess strong potential for applications in diverse electronic devices. In other words, FeCo microspheres with concave architecture exhibited superior EMWA performance under both thin thicknesses ( < 2.5 mm) and broad low-frequency bands ( > 2 GHz) compared with reported magnetic alloys and composites, positioning them at the forefront of EMWA materials for 5 G communication and microwave technologies (Fig. 6c, Supplementary Fig 30c and Table 8). To further clarify the performance distinction, we also evaluated the alloys in the mid-frequency range (8–12 GHz). The results indicate that neither the optimized μ″ nor the EMWA performance shows a notable enhancement, confirming the optimization strategy is primarily concentrated within the 2–8 GHz (Supplementary Fig. 44). EW absorption performance and application potential. a Two-dimensional (2D) performance diagram of the optimized FeCo microsphere. b EW absorption efficiency curves of dense (D-FeCo) and optimized FeCo (O-FeCo) microspheres. The red block at the bottom is related to Bluetooth, the blue block represents 5 G n79 and satellite communication. c Comparison of effective absorption band (EAB) and thickness between the alloys developed in this study and reported magnetic materials. d Three-dimensional radar cross-section (RCS) maps of various alloys at 4.7 GHz. e Schematic of absorber-covered metal backplate model and corresponding 2D RCS curves at different incident angles. f Comparison of RF response values before and after applying the magnetic patch, the inset image is a schematic of radar frequency testing. g Stress–strain curves and schematic of the mechanical performance of the optimized magnetic patch. h Suspension and sedimentation properties of optimized and dense FeCo, glycerol was used as the liquid in the surface dish. i Radar chart of key performance metrics for the optimized FeCo alloy and dense counterpart. Although RL curves generally represent absorption performance under normal-incidence EW waves, angular variability of external EM waves should be considered for practical applications48,49. CST simulations confirmed that the optimized FeCo film maintained high EMWA efficiency across varying incidence angles. In Fig. 6d, the dense FeCo structure exhibited strong scattering, whereas the optimized FeCo considerably suppressed scattered signals, reflecting improved energy dissipation. Polar plots (Fig. 6e) revealed a reduced radar cross section (RCS) of 20.1 dB in the main radiation direction (0°), equivalent to a 99% decrease in reflected intensity. Moreover, the optimized alloys displayed narrower radiation lobes and consistently lower RCS at other angles, confirming broadband angular attenuation and superior EMI suppression capability. To further evaluate performance in real devices, the powders were combined with Ecoflex to form a flexible patch that integrates magnetic loss and mechanical flexibility (Supplementary Fig. 45). Real-device EM radiation tests, conducted using a radar frequency (RF) measurement system. The stand-alone devices emitted high radiation levels under active signal conditions, with a peak intensity of 134.8 mW/m2 (red warning). When the magnetic patch was applied to the device, radiation dropped significantly to 34.4 mW/m2 (green safety), demonstrating its effective EM protection ability in practical environments (Fig. 6f). In addition to EM benefits, this patch exhibited excellent mechanical properties (elasticity and stretchability), including a 94.1% increase in fracture strain and a 56.7% enhancement in tensile strength compared with the dense FeCo-based counterparts, achieving rare synergy between mechanical stretchability and low density. This improvement arises from structural interlocking at concave and porous sites, which strengthens interfacial bonding—consistent with computational predictions. Similarly, structural optimization endowed the alloy with a lightweight nature due to significantly reduced density. As illustrated in Fig. 6g and Supplementary Fig. 46, this density difference between the optimized and dense FeCo alloys was visually confirmed through distinctive suspension and sedimentation behaviors in liquids, as well as a mass difference of 7.6 g under equal volume. These results collectively validate the effectiveness of active magnetic region engineering in designing lightweight, high-performance materials with low density, strong magnetic response, broadband absorption and superior tensile properties (Fig. 6h). In summary, this study presents a concave structural design strategy to overcome the long-standing trade-off between magnetism and density in rigid metallic systems. By integrating micromagnetic simulations with ML, we accelerated the discovery of high-activity magnetic architectures. Stress asymmetry, manipulated during droplet evaporation, enabled the synthesis of rigid magnetic alloy microspheres with controllable surface concavities. This curvature-induced design effectively reshaped the magnetic domain configuration and redistributed magnetic energy, as validated by micromagnetic modeling and off-axis electron holography. The optimized FeCo alloys achieved high saturation magnetization (236.6 emu/g) and superior magnetic responsiveness at low density (0.92 g/cm3). Moreover, the fabricated magnetic patches exhibited enhanced tensile performance. Although this study focused on magnetic alloys, this synthesis strategy can be extended to other rigid metals such as Cu, Mn, and Al, significantly broadening its potential scope of applications. We believe that this work offers a viable pathway for activity design and density optimization, establishing a paradigm for lightweight multifunctional materials with promising applications in energy utilization, water treatment, and drug delivery. Method Materials and chemicals Aluminum (III) nitrate nonahydrate (Al(NO3)3·9H2O), Ferric(III) nitrate nonahydrate (Fe(NO3)3·9H2O), cobalt(II) nitrate hexahydrate(Co(NO3)2·6H2O), nickel(II) nitrate hexahydrate (Ni(NO3)2·6H2O), copper(II) nitrate trihydrate (Cu(NO3)2·3H2O), Manganese(II) nitrate tetrahydrate (Mn(NO3)2·4H2O) were purchased from Shanghai Aladdin Reagent Co., Ltd. Citric acid (C6H8O7·H2O) and ethylene glycol (C2H6O2) were obtained from Sinopharm Chemical Reagent Co., Ltd. DI water were used from a Milli-Q system. All reagents could be used without further purification. Synthesis of concave microspheres Metal oxide microspheres were prepared using a spray drying method, followed by high-temperature reduction to obtain alloy microspheres. Taking the Fe₅Co₄Ni alloy as an example, 10 mmol of Fe (NO3)3·9H2O, 8 mmol of Co (NO₃)2·6H₂O, and 2 mmol of Ni (NO₃)2·6H₂O were dissolved in 100 mL of deionized water. Then, Then, 10 mmol of citric acid and an appropriate amount of ethylene glycol were added under thorough stirring to form a homogeneous solution. Citric acid was deliberately selected primarily for its strong chelating ability and excellent molecular-level mixing with metal precursors, ensuring a highly homogeneous precursor distribution and reproducibility during structural evolution. In contrast, polymeric carbon sources (e.g., PVP or sucrose) tend to produce thicker shells or porous morphologies, reducing structural controllability and reproducibility. The resulting solution was introduced into a spray dryer (Buchi 290) with a feed rate of 3 mL/min, an inlet temperature of 220 °C, and an outlet air temperature maintained at 100°C. The obtained precursor powder was calcined in air at a heating rate of 2 °C/min to 400 °C and held for 3 h to remove the carbon source and form stable oxide microspheres. Finally, the alloy microspheres were obtained by reduction in a H2/Ar atmosphere, heating at 1 °C/min to 450 °C and holding for 1 h. Fe7Co3 and Co3Ni1 alloys were synthesized using the same procedure, with the metal salt ratios adjusted according to their elemental compositions. Different materials were subjected to appropriate calcination atmospheres and temperatures according to their properties: FeCoNiMn and FeCoNiMnCu were reduced and calcined under a H2/Ar atmosphere at 600°C, Al₂O₃ was calcined in air at 1000 °C, and Cu was reduced and calcined at 400 °C. Characterization The crystal structure could be characterized by Power X-ray diffraction (XRD) measurements with Ni-filtered Cu Kα radiation. The morphology were observed by a Field emission scanning electron microscope (FESEM, Hitachi S-8600, Japan) and a transmission electron microscope (TEM, JEOLJEM-2100F). The EDS spectra and elemental maps were obtained using the dual EDS detectors on the Hitachi S-8600. The high-resolution transmission electron microscopy (HRTEM) and off-axis electron holography were equipped with TEM (JEOL JEM-2100F, 200 kV). Micromeritics Tristar 3020 analyzer was used to measure N2 absorption-desorption isotherm plots and specific surface area. The hysteresis loop data was acquired by the superconducting quantum interference device (SQUID VSM, Quantum Design Company of the USA). X-ray photoelectron spectroscopy (XPS) with a monochromatic X-ray source (Al Kα, hν = 1486.6 eV) was conducted to study the electment valence. Surface topography was characterized by atomic force microscopy (AFM) using an Asylum Research Jupiter XR system (Oxford Instruments). 3D topography was performed using a focused ion beam–scanning electron microscope (FIB–SEM, Helios 5 UX, Thermo Fisher Scientific), and the acquired serial images were reconstructed and analyzed using Avizo 2025.1 software. Micromagnetic simulation Alloy spheres with varying magnetic structures were simulated using MuMax3. The geometric and magnetic parameters were determined based on literature data and prior simulation results16,49–51. The specific magnetic parameters are listed in Supplementary Table 1. To account for the non-uniform magnetic distribution observed in experiments—arising from sample defects or grain boundaries—non-intrinsic damping coefficient was incorporated to improve the accuracy of simulation results. In addition, the exchange length (lex) was calculated using the standard formula according the magnetic parameters: The constructed grid size should be smaller than lex, so the size is set to 5*5*5 nm. A high-frequency magnetic field was applied to investigate the motion of magnetic domains and the corresponding energy distribution. The simulations of magnetic domain evaluation in this study were performed using the same material system as in the experiments (Fe₇Co₃), ensuring that geometric morphology serves as the sole variable influencing magnetic-domain evolution. Statistical method for energy dissipation distribution We sequentially read the magnetization vectors from each frame of the Mumax3 binary OVF outputs. For each cell, the rotation angle θi between successive frames was calculated, and local energy dissipation was defined as di=mi*sinθi, where mi is the magnetization magnitude in the cell, summing di across all cells and time steps yields the total dissipation Dtotal. For a series of angle thresholds θth, we then count the number of cells whose rotation angles exceed each threshold and sum their corresponding di. By normalizing these sums by Dtotal, we obtain the dissipation ratio as a function of the rotation-angle threshold. In a word, this method identifies active magnetic regions based on their relative energy dissipation, which can be used as the basis for determining rotation-angle thresholds. Machine learning A dataset containing 630 samples was generated from micromagnetic simulations to investigate the structure–property relationships governing magnetic response. To obtain a more robust and generalizable performance evaluation, we adopted an independent test set (20%), and five-fold cross-validation was conducted on the remaining data (64% training, 16% validation per fold) to assess the model’s overall behavior across different data partitions. Four categories of physically interpretable descriptors were used as input features: porosity, surface concavity, elemental composition, and atomic-ratio parameters. The target outputs were the peak value and minimum value of the imaginary part of permeability within the 2–8 GHz frequency range. For consistency across algorithms, all input features were standardized prior to model training. A correlation-based feature selection procedure was then applied to remove highly collinear variables and reduce overfitting. Five supervised regression algorithms were evaluated, including Random Forest (RF), K-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), Multi-Layer Perceptron (MLP), and Support Vector Machine (SVM). Model performance was assessed using the coefficient of determination (R²) and root-mean-square error (RMSE). The test set was held out before model optimization, and all hyperparameter tuning and cross-validation were conducted only on the training set to avoid data leakage. The final model was selected based on its predictive accuracy on the independent test set, together with its cross-validation performance and stability. To improve interpretability and physical relevance, domain knowledge was incorporated during descriptor selection and post hoc analysis. Feature contributions were further quantified using Shapley Additive Explanations (SHAP), which enabled identification of the structural parameters most strongly governing magnetic permeability. Statistical method of concavity coverage and roughness Considering that the concavities are already formed at the precursor oxide stage, we used the oxide microspheres to more reliably evaluate surface concavities. The samples were classified into five categories—smooth, low, moderate, high, and extreme—according to the area coverage. For quantifying the concavity coverage, the microspheres were approximated as circular projections, and each concavity was treated as a local circular pit (Supplementary Figs. 33–36). Based on the microsphere diameter D and the diameters di of individual concavities extracted from SEM images, we estimated the projected coverage fraction using the formula Σi(di/D)2. Although the concavities may exhibit certain geometric irregularities, this method remains a reasonable first-order approximation that captures the relative variation in the extent of surface concavities among different samples. In addition, the roughness parameter Sa was defined as the arithmetic mean deviation of surface heights relative to the average level, namely Sa=1N∑i=1N∣zi−z¯∣, where N is the total number of sampling points within the scanning region, zi is the measured surface height, and z¯ is the mean surface height. In practice, the Sa represents the statistically averaged roughness of the entire scanned surface (Supplementary Figs. 38–40). Statistical method of curvature distribution A representative microsphere with high concavity coverage was selected for reconstruction. From the FIB-based 3D tomography models from different orientations (Supplementary Fig. 41 and Supplementary Movie 2), the local concave features were approximated as spherical caps. For each concavity, the mouth diameter Dc and depth h were measured from the reconstructed cross-sections, and the effective curvature radius was estimated as R=((Dc/2)2+h2)/(2h). The corresponding curvature magnitude was then defined as ∣κ∣=1/R (Supplementary Table 7) Radar cross-section calculations CST STUDIO SUITE 2018 was employed to simulate the radar cross-section (RCS) of a rectangular thin perfect electric conductor (PEC) substrate (100 × 100 × 0.1 m3) coated with either dense FeCo microspheres or optimized alloy microspheres (FeCo, CoNi, FeCoNi), each with a coating thickness of 2.2 mm. The incident angle of the electromagnetic wave was varied from − 90° to 90°, and the RCS values were calculated using the following equation: Where S is the area of simulation model, λ represents the wavelength of electromagnetic wave, Es is the electric filed intensity, and Ei is the incident wave. Ansys Fluent simulation Ansys 19.0 was used to simulate the evaporation of droplets under high temperature conditions. The pressure-based solver was selected with a transient formulation, and the velocity formulation was set to absolute. The model setup included multiphase (volume of fluid), energy and laminar. The multiphase system consisted of air, water, and ethylene glycol. The saturation temperatures for water and ethylene glycol in the air phase were set to 373 K and 470 K, respectively. The surface tension coefficients of water and ethylene glycol were defined as 0.0727 N/m and 0.0484 N/m, respectively. The ambient temperature was set to 500 K, while the initial temperature of the droplets was 300 K. The volume ratio of ethylene glycol to water in the droplets was varied between 0 and 10. A time step size of 1 × 10−7 s was used, with a total of 1000 time steps. Electromagnetic wave absorption performance measurement The electromagnetic parameters and absorption performance were tested by the N5230C vector network analyzer. The RL values can be obtained according to the transmission line theory as follows: Where εr and μr are the complex permittivity and permeability of the composite. f is frequency, c represents the velocity of waves, d represents the thickness of absorbers, Z means the normalized input impedance. The samples (70 wt%) were mixed with molten paraffin evenly and pressed into a hollow coaxial ring with 7.0 mm outer diameter and 3.0 mm inner diameter. The microwave absorption rate (%) was calculated from by employing the obtained reflection loss in dB used following equation: Electromagnetic radiation detection experiment A mobile phone equipped with a communication card was used to receive and transmit signals. FeCo composite elastic patches with 2.0 mm thickness were fabricated using Ecoflex. The strength of the radiation emitted by the mobile phone was evaluated based on the dynamic radar frequency (RF) value. The recorder was positioned 3 cm away from the phone. Upon receiving an incoming call, the dynamic RF values were compared between conditions with and without the magnetic patch. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary File Supplementary Movie 1 Supplementary Movie 2 Reporting Summary Transparent Peer Review file Acknowledgements We acknowledge the Thermo Fisher Scientific for their support in three-dimensional tomography. Author contributions R.C. and W.Y. conceived and designed the research. C.Z. carried out the micromagnetic simulations, experiments, analyzed the data and wrote the manuscript. L.R. analyzed the data and wrote the manuscript. L.W. carried out the experiments and analyzed the data. R. Z. conducted the ML calculation. J.Q. performed the TEM test and electron holography. M.W., K.P., C.L., and H.C. help with the data analysis. All authors commented on the results and contributed to manuscript revisions. Peer review Peer review information Nature Communications thanks Zhihui Zeng, Guangcun Shan, and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Funding This work was supported by the National Natural Science Foundation of China (No. 12327804, 52231007, T2321003, 22088101, 22405050), the National Key Research Program of China (No. 2024YFA1208902, 2024YFA1408000, 2025YFA1614800), the Science and Technology Commission of Shanghai Municipality (No. 24ZR1406400), and the Shanghai Municipal Education Commission (No.24KXZNA06). Data availability All data necessary to evaluate the conclusions of this study are included in the manuscript or the Supplementary Information. Additional data are available from the corresponding author upon request. Source data are provided in this paper. Code availability The code used in this work has been deposited in GitHub (https://github.com/19213010015-ai/Zhang-Chang) and archived in Zenodo to provide a permanent and citable version of the exact code used in this study. The source code and associated datasets are publicly available in the repository “19213010015-ai/Zhang-Chang: ML-data” under DOI: 10.5281/zenodo.20135221. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Wenbin You, Email: wbyou@fudan.edu.cn. Renchao Che, Email: rcche@fudan.edu.cn. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-74017-9. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary File Supplementary Movie 1 Supplementary Movie 2 Reporting Summary Transparent Peer Review file Data Availability Statement All data necessary to evaluate the conclusions of this study are included in the manuscript or the Supplementary Information. Additional data are available from the corresponding author upon request. Source data are provided in this paper. The code used in this work has been deposited in GitHub (https://github.com/19213010015-ai/Zhang-Chang) and archived in Zenodo to provide a permanent and citable version of the exact code used in this study. The source code and associated datasets are publicly available in the repository “19213010015-ai/Zhang-Chang: ML-data” under DOI: 10.5281/zenodo.20135221.