The physics behind the platform.
MaterSynq is built on peer-reviewed computational methods in condensed matter physics and materials informatics. Here is what runs under the hood.
Machine Learning Interatomic Potentials (MLIPs)
MLIPs are neural network models trained to predict potential energy surfaces from quantum-mechanical reference data. The core idea: rather than solving the Schrödinger equation for every new geometry at O(N³) DFT cost, we train a model that can predict energy and atomic forces in milliseconds — at a fraction of the compute.
MaterSynq employs equivariant neural network architectures in the NequIP/MACE class. These architectures respect the physical symmetries of atomic systems: rotational invariance of energy, equivariance of forces under coordinate transforms. This physical inductive bias is critical for generalization to novel compositions outside the training distribution.
Training and validation protocol
Every MLIP deployed in production is trained on DFT datasets covering the relevant transition metal oxide chemistry space (Ni/Co/Mn oxides, layered R3̄m structures, spinel phases). Before deployment, each model is validated on a held-out test set from the same chemical family — we require <50 meV/atom MAE on formation energies and <0.1 eV/Å force RMSE as minimum passing criteria.
E_model ≈ E_DFT ± 50 meV/atom (held-out oxide test set)
Generalization bounds: MLIPs trained on layered oxides do not automatically transfer to phosphate or fluoride chemistries. For novel chemistry classes, MaterSynq requires additional DFT reference data before reliable deployment. We are explicit about this scope boundary in onboarding.
DFT Reference: PBE+U for Transition Metal Oxides
Density functional theory (DFT) is the workhorse of computational materials science. For transition metal oxides containing Ni, Co, and Mn — the core cathode materials — the standard generalized gradient approximation (GGA/PBE) systematically underestimates the on-site Coulomb repulsion of d-electrons, producing errors in formation energies and incorrect electronic ground states.
The DFT+U correction (Liechtenstein/Dudarev formulation) adds an empirical Hubbard U parameter to the d-orbitals of transition metals, correcting the self-interaction error. MaterSynq uses U values consistent with the Materials Project database (U=6.0 eV for Ni, U=3.32 eV for Co, U=3.9 eV for Mn) to ensure compatibility with the largest public convex hull reference.
Computational setup
All DFT single-point calculations use plane-wave basis sets with energy cutoffs ≥520 eV, k-point grids with spacing ≤0.03 Å⁻¹ (Γ-centered Monkhorst-Pack), and spin-polarized calculations with ferromagnetic initialization for transition metal systems. Geometry optimization uses the conjugate-gradient algorithm with convergence criteria of <1 meV/Å on forces.
Cost vs accuracy trade-off
Full DFT on 400 candidate structures is computationally expensive — approximately 2,000–8,000 CPU-hours depending on system size. The MaterSynq strategy is to use MLIP pre-relaxation to reduce the candidate pool to the top 20–40 structures before committing to full DFT, reducing total compute by 80–90% while retaining >95% of the stable-phase hit rate.
Phonon-based dynamic stability screening
A structure can have a low formation energy (sit near the convex hull) and yet be dynamically unstable — meaning it will spontaneously distort or decompose when perturbed by finite temperature. Formation energy alone is insufficient for cathode candidate filtering.
Phonon dispersion curves describe the collective vibrational modes of a crystal. A stable structure has all real phonon frequencies throughout the Brillouin zone. An imaginary phonon frequency — a mode whose frequency-squared is negative — signals a saddle point on the energy landscape: the structure will spontaneously distort along that mode.
Finite displacement method
MaterSynq computes phonon dispersions using the finite displacement (frozen phonon) method: small atomic displacements are applied to a supercell, and the resulting forces are used to construct the interatomic force constant (IFC) matrix. The dynamical matrix is then diagonalized at each q-point in the Brillouin zone. We use 2×2×2 supercells for layered oxides and sample 40+ q-points along high-symmetry paths.
Candidates with any imaginary phonon mode at q≠Γ are flagged as dynamically unstable and deprioritized in the final ranking, regardless of their formation energy. This filter removes ~15–30% of "thermodynamically plausible" candidates from synthesis consideration — saving significant experimental resources.
Surface energy computation methodology
The cathode-electrolyte interface is where most practical failure modes originate: SEI formation, oxygen release, transition metal dissolution, capacity fade. Surface energy determines which crystal facet will dominate the as-synthesized particle morphology and which terminations are exposed to the electrolyte.
Slab model construction
MaterSynq constructs symmetric slab models for each crystallographically distinct Miller plane (at minimum: (001), (010), (100), (110), (111) for layered oxides). Slabs are constructed with ≥15 Å vacuum gap and ≥12 Å of material layers to ensure surface-bulk separation. Both terminations of each slab are included to avoid polar surface artifacts.
Surface energy formula
γ = (E_slab − N × E_bulk) / (2A)
Where E_slab is the relaxed slab energy, N is the number of formula units, E_bulk is the bulk energy per formula unit, and A is the surface area. The factor of 2 accounts for both exposed surfaces in a symmetric slab. All energies computed with the same DFT settings as the bulk reference.
Connection to SEI and ionic transport
Low surface energy terminations predict the dominant morphology of synthesized particles. High surface energy facets, while less abundant, are often the reactive sites for SEI formation. The surface energy data produced by MaterSynq is directly usable for Wulff construction (particle morphology prediction) and as input for cathode-electrolyte interface reaction modeling.
See the methodology in action.
Walk through the platform workflow or request access to run your first campaign.