Machine learning uncovers temperature-stable oxide for tiny capacitors

Capacitors are the quiet workhorses of modern electronics, storing and releasing charge in radiofrequency modules, integrated circuits and microwave telecommunications systems. As engineers push to pack more functionality into ever smaller devices, the dielectric materials at the heart of these components face a seemingly impossible set of demands: they must store a large amount of energy, waste almost none of it, and do so consistently whether a phone sits in a pocket at room temperature or a base station bakes in the sun. A study published in Nature Electronics by Junlei Qi, Yiying Chen and colleagues now reports a material that appears to satisfy all three requirements at once, and the route to finding it says as much about the future of materials discovery as it does about the ceramic itself.

The difficulty lies in a long-standing trade-off. Ferroelectric barium titanate achieves an enormous dielectric constant above 3,000 thanks to its spontaneous polarization, but it suffers heavy losses above 0.03 from domain-wall motion and a wildly unstable capacitance of roughly 10,000 parts per million per kelvin, driven by phase transitions near its Curie temperature. Paraelectric strontium titanate is far cleaner, with a loss tangent near 0.0004, yet its permittivity of about 190 comes with a strongly negative temperature coefficient of roughly minus 3,294 ppm per kelvin. Simple silicon dioxide offers near-perfect stability and vanishing loss, but a dielectric constant of only around 4. Decades of compositional doping, cation ordering and octahedral tilting have delivered only incremental gains, because the search space of possible oxide compositions is astronomically large and traditional trial-and-error synthesis cannot cover it.

The research team, based at institutions including Tsinghua University, the Shanghai Institute of Ceramics and City University of Hong Kong, attacked the problem with a hierarchical machine learning strategy. Rather than asking an algorithm to predict a single winning compound directly, the framework first screened broad compositional space to identify promising structural families, and only then optimized local atomic features within the chosen family. The workflow ran in four steps: dataset construction and dimensionality reduction, global feature-driven high-throughput structure screening, local feature extraction and ranking, and finally experimental validation. Crucially, the descriptors were physics-informed, drawing on the Clausius–Mossotti relation for polarizability and volume, the Phillips–Van Vechten–Levine bond theory for bond energy and strength, and the global instability index for structural robustness.

The datasets were substantial by materials-informatics standards: 6,000 entries spanning 250 samples and 24 features for permittivity, and 3,888 entries from 162 samples for the temperature coefficient. After cleansing the data and pruning collinear features using Pearson correlation thresholds, the team tested six regression algorithms with hyperparameters tuned by ten-fold cross-validated grid search. Random forest regression proved best for predicting the dielectric constant, achieving a coefficient of determination above 0.80, while support vector regression with a radial basis function kernel handled the temperature coefficient, with R-squared above 0.65. The lower score for the thermal property reflects the genuine complexity of temperature-driven behaviour, including lattice expansion and polarizability dilution, but it was sufficient to power a virtual structure library of 132,651 candidate entries.

That virtual library pointed decisively toward one structural family: tungsten bronzes, in which corner-sharing metal–oxygen octahedra form five- and four-membered rings that tolerate large cation substitutions and octahedral tilting without losing phase stability. The team selected barium neodymium titanate, Ba4Nd9.33Ti18O54, for experimental validation because it combined a low intrinsic loss near 0.0001 at 1 megahertz, a moderate permittivity of 85 and a relatively gentle temperature coefficient of about minus 70 ppm per kelvin. Correlation analysis across site-specific local features then revealed the design rules: atomic mass and electronegativity at the four-membered-ring A-prime-double-prime sites govern permittivity, while metal–oxygen bond strength and covalency at the octahedral sites control thermal stability.

Experiments confirmed the predictions with striking fidelity. Substituting neodymium with heavier, more electronegative cations raised the permittivity from 88 in the yttrium variant to 112 with lead, and 105 with bismuth. On the octahedral site, density functional theory calculations identified germanium as the dopant forming the strongest and most covalent bonds to oxygen, and measured temperature coefficients shifted progressively from minus 118 ppm per kelvin with hafnium to essentially zero with germanium. Combining 50 percent bismuth substitution with 1 percent germanium yielded the optimized composition Ba4(Nd0.5Bi0.5)9.33(Ti0.99Ge0.01)18O54, delivering a dielectric constant near 185, a loss tangent of about 0.0005 and a temperature coefficient within plus or minus 30 ppm per kelvin, stable across frequencies from 1 hertz to 1 gigahertz.

The mechanism behind this temperature insensitivity is perhaps the most scientifically compelling part of the work. Using in situ atomic-resolution scanning transmission electron microscopy, the team mapped the displacements of A-site cations at room temperature and at 500 degrees Celsius. The average displacement grew from 7.09 to 8.75 picometres, a 23 percent increase far exceeding what simple thermal expansion could produce, and the displacement vectors became randomly oriented and non-centrosymmetric, indicating a thermally induced polar state. Quasi-harmonic density functional theory simulations explained why: as the lattice expands, the germanium-substituted octahedra tilt, and the strong covalent Ge–O bonds drive the intervening bismuth ions to shift off-centre through the bridging oxygen atoms. This cooperative polar displacement offsets the usual dilution of dielectric susceptibility that heating causes, accounting for an estimated 110 percent increase in the ionic dielectric constant at 2.5 percent volumetric expansion.

Supporting spectroscopy reinforced the picture. Temperature-dependent Raman measurements showed broadening and intensification of the A-site vibrational peaks near 120 and 140 wavenumbers, signalling growing displacement disorder, alongside a redshift of the metal–oxygen stretching mode near 540 wavenumbers that points to anharmonic effects induced by Ge–O bonding. Extended X-ray absorption fine structure at the bismuth L3 edge revealed that germanium substitution broadened the distribution of Bi–O bond lengths from a range of 2.21 to 2.62 angstroms up to 2.16 to 2.90 angstroms, directly evidencing the symmetry-breaking structural disorder that accompanies the enhanced polar displacements observed in the electron microscope.

The practical payoff came in prototype single-layer chip capacitors, fabricated by tape casting a 20-micrometre dielectric film between titanium–tungsten, nickel and gold electrodes. A 35-picofarad device held its capacitance stable from 100 hertz to 10 megahertz, and a 1-picofarad chip measuring just 0.35 by 0.35 by 0.20 cubic millimetres achieved roughly an 85 percent volume reduction compared with a leading commercial counterpart while retaining the same temperature stability class. Reliability testing was equally encouraging: the ceramic withstood a direct-current breakdown field near 40 kilovolts per millimetre, retained more than 99.8 percent of its capacitance after 1,000 thermal shock cycles between minus 55 and 150 degrees Celsius, and showed negligible capacitance drift and minimal resistivity degradation over 100 hours of highly accelerated lifetime testing at 1,000 volts and 150 degrees Celsius.

Beyond the specific ceramic, the study offers a blueprint for escaping classical performance compromises in functional materials. By pairing physics-informed machine learning with atomic-scale microscopy and first-principles simulation, the researchers showed that a seemingly intractable compositional space can be navigated rationally, and they have released their dielectric database as an open-access resource for the wider community. The material itself uses conventional raw materials and mature solid-state synthesis routes, suggesting a realistic path from laboratory to factory. If the approach generalizes as the parallel experiments across other tungsten bronze systems suggest it does, the era of discovering materials by algorithm may have quietly delivered one of electronics’ most stubbornly sought components.

Subject of Research: Machine learning-guided discovery of temperature-stable high-permittivity tungsten bronze dielectric oxides for miniaturized capacitors

Article Title: Temperature-stable high-κ oxides for miniaturized capacitors

Article References: Qi, J., Chen, Y., Qin, J., Wei, B., Pan, H., Hu, T., Fu, Z., Wu, Y., Lv, R., Zhang, F., Zhang, Y., Xu, W., Zeng, J., He, S., Zhu, Z., Su, H., Yang, Z., Fu, Z., Liu, Z., … Lin, Y.-H. (2026). Temperature-stable high-κ oxides for miniaturized capacitors. Nature Electronics. https://doi.org/10.1038/s41928-026-01721-1

Image Credits: AI Generated

DOI: 10.1038/s41928-026-01721-1

Keywords: dielectric materials, high-permittivity oxides, tungsten bronze, machine learning, capacitors, temperature stability, bismuth doping, germanium doping, density functional theory, electron microscopy, chip capacitors, materials discovery