Sulfur-modified iron-based biochar via feature-sparsified cascade transfer learning: from interfacial structure...

Iron-based biochar is highly effective for phosphorus recovery; however, high iron loading often leads to the aggregation of active sites, thereby limiting its performance. While sulfur doping can regulate the interface to enhance iron dispersion, traditional experimental optimization is hindered by the non-linear coupling of synthesis parameters and the scarcity of public data. To address these challenges, this study constructed a feature-sparsified cascaded transfer learning (FS-CTL) framework. By employing the relatively data-abundant Cr(VI) removal task as the source domain, L₁ regularization was introduced during the transfer fine-tuning stage to effectively attenuate feature weights related to the reduction in the source domain. Based on this framework, the optimal precursor ratios were determined as S/C = 0.32 and Fe/C = 0.42. Batch adsorption experiments demonstrated that the adsorbent prepared via this scheme achieved a maximum adsorption capacity of 78.6 mg·g^-1, surpassing the capacities of the sulfur-free and high-sulfur control groups by 20.5% and 14.1%, respectively, while maintaining stable removal efficiency across a pH range of 2.0-9.0. Characterization results indicated that appropriate sulfur doping inhibited the aggregation and crystalline growth of iron species. Density functional theory (DFT) calculations further revealed that sulfur atoms increased the binding energy of iron on the carbon substrate through orbital hybridization, thereby restricting surface migration. This study provides a mechanistic basis for understanding the structural regulation function of sulfur in iron-based biochar systems and demonstrates the applicability of feature-sparsified transfer learning in cross-task materials design.