<p>Alloying anodes offer high capacities for lithium-ion batteries but suffer from severe structure degradation during cycling due to irreversible atomic rearrangements upon delithiation. Resolving this requires an atomic-scale understanding of structure recovering dynamics, hindered by limitations in characterization and simulation techniques. We employ deep potential molecular dynamics (DPMD), a machine-learning approach enabling large-scale simulations with near-density functional theory accuracy, to model phosphorus anode evolution. DPMD reveals that post-delithiation, isolated P atoms form disordered, loosely packed structures with low atomic density (∼34 atoms nm⁻³), significantly below the theoretical value (46 atoms nm⁻³), impeding recovery. Crucially, DPMD identifies NiP₂ crystal surfaces as nucleation sites, inducing ordered and dense P atom packing (∼42 atoms nm⁻³). Guided by these simulations, we synthesized a crystalline NiP₂-amorphous P hybrid anode. This anode exhibits reduced volume expansion (20.2% decrease), near-complete structural recovery (113.7% initial volume), and enhanced cycling stability (68.8% capacity retention after 600 cycles at 6 A g⁻¹). This work establishes a paradigm for controlling structural evolution in alloy anodes via transformative mediators.</p>