Hybrid deep learning assisted understanding of Na-Li Co-doping effects on the structural and superconducting properties of Bi-2212 ceramics

GÜRSUL M., KARAÇORA F., ULUOCAK İ., ÖZÇELİK B.

Journal of Alloys and Compounds, vol.1061, 2026 (SCI-Expanded, Scopus)

  • Publication Type: Article / Article
  • Volume: 1061
  • Publication Date: 2026
  • Doi Number: 10.1016/j.jallcom.2026.187460
  • Journal Name: Journal of Alloys and Compounds
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Chemical Abstracts Core, Chimica, Compendex, INSPEC, Public Affairs Index
  • Keywords: Bi-2212 superconductors, CNN-LSTM model, Flux pinning, Optimization
  • Hakkari University Affiliated: Yes

Abstract

The effects of Na-Li co-doping on the superconducting and magnetic properties of Bi2Sr2Ca1-xNaxCu2-yLiyO8 (x = 0.00, 0.05, 0.075, 0.10; y = 0.00, 0.05, 0.10, 0.15 and 0.20) ceramics were systematically investigated using a combined experimental and machine-learning approach. Magnetoresistance and magnetic hysteresis measurements were carried out on sixteen compositions to determine the offset critical temperature (Tcoffset) and remanent magnetization (Mr), which serve as key indicators of intergranular coupling and flux pinning. Experimentally, Tcoffset shows strong magnetic-field dependence due to enhanced vortex dynamics in intergranular regions, while Mr decreases with increasing temperature as a result of thermally activated vortex motion. To model the complex, nonlinear relationships among composition, temperature, magnetic field, and superconducting response, deep-learning regression models were developed and optimized via Bayesian optimization. A hybrid CNN-LSTM architecture achieved the highest predictive accuracy for both Tcoffset and Mr, outperforming standalone recurrent models. Sensitivity analysis based on permutation importance identified Na and Li concentrations as the dominant parameters governing both responses, in agreement with experimental trends. Furthermore, machine-learning generated contour maps revealed optimal co-doping regions that enhance intergranular coupling and flux pinning while mitigating magnetic-field suppression. These results demonstrate that machine-learning-assisted analysis provides an effective framework for interpreting and optimizing the superconducting performance of complex oxide ceramics