Composition-Only Machine-Learning Benchmark for Three Dielectric Properties of Single-Phase ULTCC Ceramics

This study establishes a composition-only machine-learning benchmark for three dielectric properties of single-phase ultra-low-temperature co-fired ceramics (ULTCCs): relative permittivity 𝜀 r , quality factor Q × f, and temperature coefficient of resonant frequency 𝜏 f . ULTCCs were selected as a processing-constrained subset of LTCC materials in which reduced-temperature densification must be achieved without sacrificing microwave dielectric performance. The three properties were modeled independently as separate single-output regression tasks using 145 Magpie descriptors derived from chemical formulas. Bayesian ridge regression, support vector regression, Gaussian process regression, random forest, and XGBoost were evaluated under a common modeling framework, with genetic algorithms used only for hyperparameter optimization. Under the current test-set evaluation, the best R 2 values were 0.8957 for 𝜀 r , 0.4689 for Q × f, and 0.0344 for 𝜏 f . These results indicate that composition-derived descriptors provide comparatively useful predictive information for 𝜀 r , limited information for Q × f, and insufficient information for reliable prediction of 𝜏 f . The framework should therefore be regarded as a property-dependent composition-only benchmark rather than as a generally reliable predictive tool for all three dielectric properties. Improving the prediction of Q × f and 𝜏 f will require additional information related to ionic polarizability, crystal structure, processing conditions, density, and microstructure.