Cobalt: Offline Contextual Bandit Learning

• The paper introduces Cobalt, a novel framework that integrates offline trajectories into contextual bandit learning to improve policy evaluation.
• It employs rigorous theoretical analysis and empirical benchmarks to demonstrate enhanced learning efficiency and reduced exploration requirements.
• The study underscores practical applications in personalized recommendation systems and adaptive decision-making, offering actionable insights for deployment.

Lithium-excess layered cathodes, commonly formulated as $(1-x)\,\mathrm{LiMO_2} + x\,\mathrm{Li_2MnO_3}$ with $M=$ Mn, Ni, Co, etc., constitute a major class of high-capacity intercalation materials for rechargeable lithium-ion batteries. This family exploits combined cationic and anionic redox, phase-engineered stacking, and tunable defect chemistry to achieve reversible capacities significantly exceeding those of conventional layered oxides. The scope, design logic, atomistic transport mechanisms, and electrochemical performance of Li-excess layered cathodes have been significantly elucidated through advanced first-principles calculations, machine-learning frameworks, and operando characterization.

1. Structural Motifs and Phase Intergrowths

Li-excess layered oxides are defined as composite materials containing two canonical building blocks: the R$\bar{3}$m-type (classical layered) $\mathrm{LiMO_2}$ and the monoclinic C2/m-type $\mathrm{Li_2MnO_3}$. The prototypical composition is $(1-x)\,\mathrm{LiMO_2} + x\,\mathrm{Li_2MnO_3}$, where $x$ typically ranges from 0.1 to 0.7 (Lee et al., 2 Feb 2026). In the resulting crystals, pure Li layers, mixed transition-metal (TM)/Li layers, and various cation-site orderings coexist. The Li$_2$MnO$_3$ end member (i.e., $x=1$) is structurally characterized by the ordered rocksalt superstructure, with Li and Mn occupying distinct octahedral Wyckoff sites: 2$b$ (mixed Li/Mn-O slabs), 4$h$ (pure Li layers), and 2$c$ (interlayer Li) (Lee et al., 2 Feb 2026, Hoang, 2014).

In practice, these oxides form stacking intergrowths and superstructures detectable via extra Bragg reflections and local distortions. Operando 3D Bragg coherent diffractive imaging (BCDI) directly reveals superstructure collapse, TM/Li disordering, and stacking faults during electrochemical cycling (Singer et al., 2017).

2. Defect Chemistry and Site Occupancy

The defect landscape is highly sensitive to composition and synthesis conditions. In Li$_{1+\delta}$Co$_{1-\delta}$O$_2$, excess Li stabilizes negatively charged Li antisites (Li$^+$ on Co$^{3+}$ sites) and small hole polarons (Co$^{4+}$ at Co$^{3+}$ sites), with oxygen vacancies being energetically suppressed under Li-rich, Co-poor environments. The defect chemistry obeys:

$$\mathrm{LiCoO_2} + \mathrm{Li_2O} \rightarrow \mathrm{Li}_{1+\delta}\mathrm{Co}_{1-\delta} \mathrm{O}_2,\quad \delta \simeq [\,\text{Li}_\text{Co}'\,]=2[\textrm{h}^\bullet]$$

The binding energy of the Li$_\text{Co}'$ + 2 h$^\bullet$ complex reaches $\approx 2.23$ eV, and concentrations exceeding $10^{20}$ cm$^{-3}$ are thermally accessible at typical synthesis temperatures. Charge transport is mediated primarily by small polaron hopping (with migration barriers as low as 0.10 eV for h$^\bullet$) and Li vacancies (barriers of 0.18–0.70 eV), the latter enhanced under Li-excess conditions owing to aggregation into divacancy clusters (Hoang et al., 2014).

In Li$_2$MnO$_3$, dominant intrinsic defects include Mn$^{2+}$ antisites and Li vacancies. The Mn$_\text{Li}^+$ formation energy can be raised above 2 eV by synthesizing under Li-rich, Mn-poor, and O-rich conditions, reducing antisite disorder. Delithiation is coupled to O$^{2-} \to$ O$^-$ oxidation and bound O-hole polaron ($\eta_\mathrm{O}^+$) formation, with poor electronic conductivity at low Li-vacancy concentrations (Hoang, 2014).

3. Li-Ion and Electronic Transport Mechanisms

Li-ion mobility in Li-excess layered oxides integrates both monovacancy and divacancy migration, strongly modulated by structural disorder, dynamical correlations, and short-range order. In paramagnetic Li$_2$MnO$_3$, recent DFT+$U$+DMFT+NEB studies classify six symmetry-inequivalent Li$^+$ hops: four intralayer and two interlayer pathways. Dynamical mean field theory (DMFT) dramatically lowers the lowest activation energies relative to static DFT+$U$, yielding:

• Shortest-range hop (4$h \to T^{\mathrm{Li}_{2b}\to4h}$): $E_a = 0.18$ eV—quantitatively matching $\mu^+$SR measured values.
• Long-range bottleneck (4$h \to T^{\mathrm{Li}_{2b}\to2c}$): $E_a = 0.50$ eV—aligned with AC-impedance experiments.

Static Hubbard $U$ corrections alone are insufficient; dynamical electronic correlations and finite-temperature screening substantially lower migration barriers, especially for short-range hops (Lee et al., 2 Feb 2026).

Electronic conduction occurs via small-polaron hopping, with carriers (hole or electron polarons) localized on transition-metal or oxygen sites. In Li$_2$MnO$_3$, $\eta_\mathrm{O}^+$ transport is contingent on Li-vacancy percolation, limiting conductivity at low delithiation (Hoang, 2014). No bandlike carriers are present; doping n- or p-type is unachievable and Fermi energy is pinned near midgap (Hoang et al., 2014, Hoang, 2014).

4. Influence of Disorder and Short-Range Ordering

Cation disorder and short-range Li clustering directly impact the percolation of low-barrier Li$^+$ migration channels in both layered and disordered rocksalt (DRX) variants. High-throughput DFT and machine-learning (ML) studies introduce two effective descriptors:

• Phase-stability $F = E_\text{hull} - T_0 S_\text{mixing}$, reflecting equilibrium likelihood of an ordering motif.
• SRO descriptor $$\Delta E_\mathrm{SRO} = E_\text{spinel-like} - E_{\gamma\text{-LiFeO}_2}$$, quantifying Li$_4$ cluster preference (key for fast diffusion).

The DRX and Li$_4$-clustering propensities are element-specific and can be tuned combinatorially (Sc, Ti, V, Mn, Fe, Co, Ni, Sn, etc., promote DRX; Ru, Ir, Mo, Mn, Co, V favor favorable SRO) (Liu et al., 14 Jun 2025). Statistical analysis across $>$6,000 compositions maps elemental contributions to DRX/Li$_4$-cluster stabilization, guiding rational dopant and composition selection.

In Li$_2$MTiO$_4$ (M = V, Cr, Mn, Fe, Co, Ni), DFT+U calculations reveal near-degeneracy between layered and disordered polymorphs ($|\Delta E| \lesssim 0.1$ eV/f.u.), enabling three-dimensional percolation without significant energy penalties. This near degeneracy vanishes in the Na analogues, where size mismatch locks in layered order (Yamauchi et al., 2021).

5. Anionic Redox, Dislocation Network Formation, and Voltage Fade

Anionic (oxygen) redox is central to the high capacity of Li-excess layered cathodes. Delithiation triggers O$^{2-} \to$ O$^-$ and O$^-$ to O$_2^{2-}$ conversion, and, at high state of charge, O$_2$ evolution. The nucleation of partial dislocations, as visualized by BCDI, disrupts the long-range TM/Li stacking order, generating regions of local disorder that facilitate oxygen redox but also cause irreversible cation mixing (Singer et al., 2017). The resulting dislocation density in Li-rich layered oxide (LRLO) reaches $1\times10^{10}$ cm$^{-2}$ at 4.4 V, an order of magnitude higher than classical layered oxides.

Pipe diffusion of oxygen along dislocation cores enables high capacities (exceeding 300 mAh/g), but the loss of superstructure with associated cation disorder leads to voltage fade. This fade is fundamentally reversible; high-temperature annealing ($T>150^\circ$C, O$_2$ atmosphere) reestablishes superstructure peaks and restores both voltage profile and reversible capacity (Singer et al., 2017).

6. Design Principles Informed by First-Principles and ML

The design and optimization of Li-excess layered cathodes utilize the following principles:

• Match transition-metal ionic radii to Li$^+$ to favor disordered topologies and percolating Li channels (Yamauchi et al., 2021).
• Favor shallow formation enthalpies ($|\Delta H_f| \lesssim 0.2$ eV) for voltage flatness and solid-solution behavior (Yamauchi et al., 2021).
• Combine phase-stability and SRO descriptors to select compositions with both thermodynamic equilibrium stability and favorable short-range Li$_4$ clustering (essential for low-barrier, fast Li$^+$ conduction) (Liu et al., 14 Jun 2025).
• Exploit ML models parametrized by ionic radii differences to predict the relative stability of layered versus disordered phases (Yamauchi et al., 2021, Liu et al., 14 Jun 2025).
• Control synthetic conditions (e.g., Li-rich, TM-poor, O-rich environment) to suppress antisite disorder and optimize defect populations (Hoang et al., 2014, Hoang, 2014).

Practical implementation is exemplified by DRX LiCr$_{0.75}$Fe$_{0.25}$O$_2$ and its 20% Li-excess variant, where the SRO descriptor accurately predicted percolating Li$_4$ environments and observed high capacities (234 and 320 mAh/g, respectively) (Liu et al., 14 Jun 2025).

7. Outlook and Ongoing Challenges

Li-excess layered cathodes embody the fundamental interplay among crystal chemistry, defect physics, correlated electronic structure, and mesoscopic disorder. While DFT+$U$ and hybrid functional calculations delineate static barrier landscapes and defect energetics, only finite-temperature DMFT frameworks consistently unify short- and long-range transport with experiment—demonstrating that dynamical correlations are essential to realistic modeling (Lee et al., 2 Feb 2026).

One persistent challenge is stabilizing high anionic redox capacity without inducing irreversible cation disorder and voltage fade. Tailored synthesis, nanostructuring, and strategic cation substitution present continuing pathways to address these issues (Hoang, 2014, Liu et al., 14 Jun 2025). The integration of high-throughput DFT, ML-guided screening, and operando experimentation will continue to define the compositional and structural boundaries of next-generation Li-excess cathodes.