FISTA-Optimized Multi-Kernel Sparse Regression for Predicting the Ionic Conductivity of Solid-State Electrolytes in Lithium Batteries
The ionic conductivity of lithium solid electrolytes is measured one material at a time by impedance spectroscopy. Predicting it from existing measurements requires not only accuracy but also an indication of which groups of descriptors carry the signal. This paper proposes multi-kernel sparse regression (MKSR). A dictionary of base kernels is built on groups of descriptors, and spectral feature maps make the Euclidean norm of each coefficient block equal to the reproducing-kernel Hilbert-space norm of the corresponding component, so that a sparse-group penalty selects descriptor groups and kernel scales. For this objective we give the gradient Lipschitz constant, its insensitivity to the spectral truncation threshold, the weighted sparse-group proximal operator and a blockwise characterisation of the regularisation level at which the solution vanishes, and we minimise the objective by the fast iterative shrinkage-thresholding algorithm (FISTA) with adaptive restart; a corresponding Lipschitz bound also holds for a censored (Tobit-type) loss. On 599 measurements from Open solid Battery Electrolytes with Li: an eXperimental dataset (OBELiX) and 820 Liverpool measurements, under cross-validation grouped by reduced composition, MKSR attains mean absolute errors of 0.985 and 0.990 in log10 conductivity, lower than equal-weight multiple-kernel ridge regression (1.044 and 1.039) also under a corrected resampled t-test, and comparable to the best of ten baselines, including extreme gradient boosting (XGBoost). FISTA with adaptive restart needs fewer iterations than the iterative shrinkage-thresholding algorithm (ISTA), and its speed is compared with coordinate-descent and quasi-Newton solvers. MKSR activates on average 4.0 of 37 kernels on OBELiX and 2.9 of 31 on Liverpool, concentrated on small-bandwidth Laplacian kernels on lattice and stoichiometry descriptors. This selection is stable under changes of the tuning grid, the spectral threshold and the common-bound treatments of censored records examined here, while leave-one-structural-family-out tests show the limits of extrapolation to unseen families.
Authors
- Tingzhang Yang (ORCID: https://orcid.org/0000-0002-6271-4944)
- Xiaokun Du
Institutions
- Chongqing Technology and Business University (CN)
- Jilin University (CN)
Publication Details
- Journal
- Mathematics
- Published
- 2026-11-19
- DOI
- https://doi.org/10.3390/math14224266
- Primary Topic
- Machine Learning in Materials Science
- Type
- article
- Field-Weighted Citation Impact
- 0.00