The phrase "crystal structure prediction" covers a lot of ground. In the most general sense, it describes the computational problem of finding, given only a chemical composition, all crystalline phases that are energetically accessible under given conditions. In the more applied sense relevant to cathode materials, it means something narrower: given a target chemistry and crystal prototype, enumerate candidate arrangements, relax them, and rank them by stability. This article describes the second problem — the practical workflow, not the fully general combinatorial challenge.
A materials scientist working on a new Ni-rich layered oxide composition typically starts with a hypothesis: "I want to substitute a fraction of Co with Al at the 3a Wyckoff site in the R3̄m prototype. What stoichiometries are accessible, and which are stable enough to synthesize?" Turning that hypothesis into a synthesis-ready candidate list is a four-step process.
Step 1: Prototype enumeration
The starting point is a crystal prototype — a Wyckoff site assignment that defines the topology of a crystal structure without specifying the exact atomic species or lattice parameters. For layered oxide cathodes, the dominant prototype is R3̄m (rhombohedral, space group #166 in the International Tables), which describes the stacking sequence of LiO₂ and TMO₂ layers characteristic of LiCoO₂ and the NMC family.
Within this prototype, a substitution question — "how many distinct ways can I arrange Ni, Co, Mn, and Al across the 3b transition metal sites?" — is a combinatorial enumeration problem. The AFLOW-POCC (partial occupancy) framework and related tools generate all symmetry-inequivalent configurations for a given composition range, respecting the space group symmetry of the prototype. For a ternary substitution in a 24-atom supercell, this typically produces 50–400 symmetry-distinct structures.
The enumeration step is computationally cheap — milliseconds to seconds. Its importance is in completeness: manual structure building misses inequivalent configurations that turn out to be the lowest-energy polymorphs. The computational enumeration is not creative; it is exhaustive.
Step 2: MLIP-accelerated relaxation
Four hundred DFT geometry optimizations, each requiring 2–4 hours of wall clock time at standard VASP settings, would take months on a typical HPC allocation. The MLIP pre-relaxation step reduces this to hours by using a physics-informed neural potential to perform the geometry optimization in place of DFT.
The MLIP relaxation is not approximate in the sense of being sloppy — it is solving the same problem (find the energy minimum of the atomic coordinates) but with a surrogate energy function that is 100–1000× cheaper to evaluate. The accuracy of the relaxed geometry depends on the MLIP's ability to reproduce the DFT potential energy surface near the minimum. For well-trained potentials on in-distribution chemistries, the relaxed MLIP geometry typically deviates from the DFT-relaxed geometry by less than 0.05 Å in bond lengths and less than 2° in bond angles.
After MLIP relaxation, a geometric stability filter removes structures that did not converge to a physically reasonable minimum: excessive bond lengths, unphysical coordination environments, or structures where the algorithm ran into a flat region of the potential energy surface. This filter typically removes 5–15% of the initial enumeration as non-physical.
The output of the MLIP stage is a set of pre-relaxed structures ranked by MLIP energy, with the clearly unstable candidates filtered out. The top 20–40 structures — those within roughly 200 meV/atom of the apparent MLIP-predicted ground state — are passed to DFT validation.
Step 3: DFT single-point validation
DFT single-point calculations (fixed geometry, energy and electronic structure only) are applied to each MLIP-pre-relaxed structure. For transition metal oxides in the Ni/Co/Mn family, the appropriate functional is GGA+U (PBE with Hubbard U corrections on the d-orbitals of transition metals). The U values must be consistent with the convex hull reference database being used — MaterSynq uses the Materials Project U parameter set (U_Ni = 6.0 eV, U_Co = 3.32 eV, U_Mn = 3.9 eV) to ensure hull distances are computed on a consistent energy scale.
A DFT single-point calculation on a relaxed 24-atom structure takes approximately 30–90 minutes, compared to 2–4 hours for a full geometry optimization. Since the MLIP has already produced a near-converged geometry, the DFT single-point captures the majority of the accuracy benefit without the computational overhead of full DFT relaxation.
For structures in the top 10 of the ranking — the most likely synthesis candidates — a full DFT geometry optimization is performed to confirm that the MLIP-relaxed geometry is close to the DFT minimum and to produce surface-energy-quality slab models for the final output package.
Step 4: Stability ranking and output packaging
Formation energies from the DFT single-points are referenced against the Materials Project convex hull for the same chemical system. The hull distance ΔE_hull (in meV/atom above the convex hull) is the primary stability metric: 0 meV/atom means thermodynamically stable, positive values indicate increasing tendency toward decomposition into competing phases.
Phonon dispersion calculations are performed on structures that pass a hull distance threshold (typically ≤50 meV/atom above the hull). The finite displacement method, using 2×2×2 supercells with 40+ q-points along high-symmetry Brillouin zone paths, identifies structures with imaginary phonon modes — dynamically unstable phases that will spontaneously distort or decompose under synthesis temperatures. These are flagged and deprioritized in the ranking regardless of their formation energy.
The final output is a ranked stability table with the following columns for each candidate: formula, space group, ΔHf (meV/atom), hull distance (meV/atom), voltage estimate (V vs Li/Li⁺), surface energy for the dominant (001) facet (J/m²), phonon stability flag, and a synthesis-readiness assessment. VASP POSCAR and CIF files are exported for all ranked candidates.
What this workflow does not provide
The workflow described here addresses thermodynamic and dynamic stability under the implicit assumption of ground-state conditions. It does not predict: electrochemical cycling stability at elevated state-of-charge, capacity fade rates over hundreds of cycles, synthesis yield from any particular synthesis route, or performance under specific electrolyte formulations. These require experiment.
The output of this workflow is a ranked candidate list for targeted experimental validation — not a substitute for that validation. The appropriate framing is: computational screening reduces the number of synthesis experiments from an undifferentiated scan of composition space to a targeted set of the highest-probability candidates. A well-executed campaign saves experimental resources; it does not replace the experiment.
For researchers evaluating whether this workflow fits their program: the minimum useful input is a defined composition range and a target property objective (stability threshold, voltage window, surface energy preference). The platform handles the rest. A typical 50-candidate Pilot campaign, from composition specification to ranked output, completes in 10–14 business days.