Research

First-principles materials discovery.

A research pipeline built around physics, not pattern-matching. Machine-learned screening at the front, density-functional perturbation theory at the back, and a discipline of predictions filed before testing — with public retrodiction holding the whole thing together.

Approach

The Polyphase Coherence framework

Our research program centers on the Polyphase Coherence framework — a theoretical model relating phonon-spectrum structure to superconducting coherence. The framework predicts which classes of materials are most likely to host high-temperature superconductivity and provides a falsifiable scoring criterion that can be evaluated from first-principles phonon calculations.

Manuscripts are in preparation. Methodology, scoring criteria, and candidate-library details are available to qualified collaborators under research agreement.

Phonon DOS DFPT Eliashberg α²F(ω) Electron-Phonon Coupling Wannier Fermi Surface Falsifiable Predictions No Back-Fitting

Pipeline

From foundation models to first-principles validation.

Foundation-Model Screening

Machine-learned interatomic potentials (MLIPs) act as fast surrogates for full quantum-mechanical calculation. Trained on first-principles data, they predict forces, energies, and phonon spectra at speeds that let us screen tens of thousands of candidate structures per week.

First-Principles Validation

Top candidates from MLIP screening are validated with full density-functional perturbation theory using Quantum ESPRESSO — the canonical method for phonon dispersion and electron-phonon coupling. Externally cited predictions are backed by DFPT, not surrogate inference alone.

Material Families

Active research spans cuprates, bismuthates, layered nickelates, layered halide systems, and intermetallics — the broad landscape where the Polyphase Coherence framework predicts the strongest signal.

Discipline

We kill our darlings.

A computational materials pipeline that only logs hits is a pipeline that overfits. We lock predictions before we run them, and we retire candidates publicly when secondary verification fails. The internal record of nulls is treated with the same rigor as the record of survivors.

MLIP screening is treated as a hypothesis generator, not as ground truth. No prediction is reported externally without independent first-principles confirmation.

Predictions on the record

Calls are filed before the validation runs — no back-fitting.

Retrodiction first

The framework is calibrated against known systems before forward predictions are made.

Public nulls

Failed predictions are retired on the record.

Member of

NVIDIA
Inception

Accepted 2026-04-02

Built on NVIDIA

Armadillo Labs is an NVIDIA Inception member. Foundation-model screening and accelerated DFPT validation run on the NVIDIA stack alongside our cloud HPC pipeline.

GPU acceleration is what lets us screen entire material families in a single batch instead of one structure at a time.

Collaborate

We work with synthesis groups, university collaborators, and capital partners.

Methodology, candidate-library details, and on-the-record prediction history are available under research agreement. If you operate at the intersection of high-temperature superconductivity, layered transition-metal oxides, or high-pressure synthesis — or you underwrite deep-tech materials — reach out.