787 oxide cathodes coloured by DSI. Scroll to fly to the eleven validation cathodes.

Structure predicts durability. Metal identity does not.

2,774 lithium-ion insertion electrodes from the Materials Project, screened through an eight-layer funnel and ranked by a three-descriptor Degradation Screening Index that is validated against 11 experimentally characterised cathodes.

See the results Source and data

Scroll

01 · Validation

Three structural descriptors rank durability better than the metal a cathode contains.

The Degradation Screening Index combines delithiation volume strain, the hull energy of the charged phase and the band gap of the discharged structure. On 11 cathodes spanning layered oxides, olivines, spinels and a Li-rich compound, it beats the single-strain descriptor and the metal-identity heuristic it replaces.

DSI against reported cycle life, 11 cathodesABCD

Spearman ρ = 0.76 (p = 0.007) for the four DSI classes; 0.82 for the continuous score. Hover a sphere for its three descriptors. Drag to orbit.

Predictor (n = 11)	Spearman ρ	p
Structure-resolved DSI, four classes	0.76	0.007
DSI, continuous score	0.82	0.002
Volume strain alone, same binning	0.55	0.08
Metal-identity heuristic	0.55	0.08

Robust to leaving any one material out (DSI ahead in 11 of 11 iterations), to removing LTO (ρ = 0.77), and to 24 meV/atom Gaussian noise on the DFT hull energies (median ρ unchanged at 0.76). Bootstrap 95 % CI [0.51, 0.93]. Weights are expert-set, not learned: nested leave-one-out weight fitting on 11 materials collapses to ρ = 0.18.

02 · One composition, three durability classes

LiCuO₂ has eight polymorphs in the Materials Project. Three are electrode end-members, and they land in three different classes.

A composition heuristic gives all three the same number. The DSI reads what each structure does when lithium leaves. The structures below are the cached Materials Project entries, drawn as 2 × 2 × 2 supercells.

C2/mground state · mp entry

C2/m, the ground state

Lowest energy, so a database search by stability picks it first. Pull the lithium out and the lattice collapses by almost half its volume; the charged phase sits 286 meV/atom above the hull.

46.6 %ΔV/V on delithiation

286meV/atom, charged phase

0.09DSI · class D

R-3m, the layered form

Layered like LiCoO₂. The volume change on delithiation is small, but the charged phase is far from stable, which the DSI penalises.

4.7 %ΔV/V on delithiation

164meV/atom, charged phase

0.46DSI · class C

Cm, lithiated from Li₁ to Li₁.₅

Its electrode entry covers lithiation at 2.9 V. Modest strain and a near-hull charged phase give the best DSI of the three.

6.6 %ΔV/V on delithiation

25meV/atom, charged phase

0.67DSI · class B

The three polymorphs on the three DSI inputsΔV/VE_hull,chargeDSI

Each bar is the value as a share of its axis maximum (50 %, 300 meV/atom, DSI 1.0); the labels carry the values. Same formula, three classes.

03 · The screen

Eight layers, from 2,774 database entries to 26 candidate compositions.

Hard filters first, scored criteria next, structure last. Every count below is regenerated by reproduce.sh from the cached Materials Project responses in the repository.

The screen as a funnel2,774 → 1,886 → 1,864 → 813 → 787 → 26

Layers 1 and 2: voltage window 2.5 to 5.0 V, capacity above 50 mAh/g (1,886). Layer 4: toxic and rare elements removed (1,864). Layer 3: oxide frameworks (813). Layers 5 to 7 score supply-chain criticality, recyclability and thermal safety without filtering. Layer 8: DSI assigned where the discharged structure has a band gap (787). Discovery zone: top-quartile composite, E_hull < 50 meV/atom, top-quartile energy density (26).

DSI = 0.50·S_vol + 0.35·S_stab + 0.15·S_bg + 0.10·[phosphate] (capped at 1)

S_vol = max(0, 1 − (ΔV/V) / 14.6 %)
S_stab = max(0, 1 − E_hull,charge / 0.190 eV/atom)
S_bg = max(0, 1 − |E_g − 2.0 eV| / 4.0 eV)

Normalisation bounds are the 95th percentiles of the 2,774-entry set. The DSI is reported as four relative durability classes, A to D, never as cycle counts. The metal-identity heuristic it replaces correlates with the structure-based strain indicator at Pearson r = 0.151 across the 569 unique formulas of the merged set.

Energy density is handled by physics, not by fitting: once each electrode is matched to the density of its own structure, the identity V × C × ρ reaches R² = 0.999 against the database values, so the screen uses it directly and keeps the learned model for incomplete records and interval estimates.

04 · Reproduce it

Every number on this page comes out of one script.

No API key is needed: the Materials Project responses are cached in the repository. The pipeline, the identifier-level merge, the DSI analysis and the figures take about two minutes on a laptop.

git clone https://github.com/Soham-ChemE/CathodeAI-Battery-Materials-Screener.git
cd CathodeAI-Battery-Materials-Screener
pip install -r requirements.txt
./reproduce.sh          # pipeline + DSI analysis + figures + docs/assets

Cite

S. Kavathekar, CathodeAI: structure-aware degradation screening of lithium-ion cathodes. Preprint, ChemRxiv, doi 10.26434/chemrxiv.15010709/v1. Manuscript under revision at Physical Chemistry Chemical Physics.

Scope and assumptions

The DSI ranks durability within the chemistry families it was validated on, with expert-set weights by design; generated compositions are screening-level priors; energy densities are theoretical values for full lithium extraction.