How It Works Platform Why Us Evidence Founder Get in Touch
Project Hephaestus

AI-Accelerated Materials Discovery

Discovering materials that don't exist yet.

UnderationAI discovers novel materials for defense and extreme-environment applications — ceramics, coatings, high-entropy alloys, and structural composites — using autonomous AI screening, first-principles DFT validation, and molecular dynamics simulation. Target property envelope in, validated candidate shortlist out.

0

Autonomous Discovery
Iterations Completed

0

Neural Network Models
Screening Every Candidate

0

Hardening Pipeline Steps
Per Candidate

0

Known Reference Materials
in Benchmark Set

Built with Quantum ESPRESSO · LAMMPS + MACE · NVIDIA CUDA · PyTorch · ALIGNN · BoTorch

The Problem

Materials discovery for defense is stuck in the last century.

Developing a new defense material takes 10–15 years and costs $50M+. Legacy ceramics, coatings, and structural alloys are decades-old technology operating near their theoretical limits. The mission can't wait for the next breakthrough to come from trial and error.

10–15 Years Per Material

From initial concept to fielded production, a new defense material takes over a decade and $50M+ in R&D. The failure rate exceeds 95%.

Unexplored Chemical Space

Multi-component ceramics, high-entropy alloys, and novel nitrides/carbides represent a vast, unexplored design space. Over 10⁶ possible compositions — too many for brute-force testing.

Evolving Mission Requirements

Hypersonic platforms, directed energy systems, and next-generation threats demand materials with properties no existing material can deliver. The pace of evolution demands faster materials innovation.

How It Works

From blank chemical space to validated material candidate.

Our platform autonomously generates, screens, validates, and hardens novel material compositions — compressing years of R&D into weeks of computation.

1

Generate Novel Compositions

Three generators — crystal structure prediction, high-entropy alloy explorer, and composition explorer — produce hundreds of novel candidates per iteration from a configurable element palette spanning ceramics, borides, carbides, nitrides, coatings, and refractory alloys.

2

Screen with 4 Neural Networks

MACE, ALIGNN, MatterSim, and MEGNet evaluate every candidate on GPU for hardness, elastic modulus, fracture toughness, and density. Thousands screened per iteration — only the most promising survive.

3

Validate with DFT Physics

GPU-accelerated Quantum ESPRESSO performs first-principles calculations: formation energy, thermodynamic stability (energy above hull), stress tensor, and equilibrium pressure. Ground-truth physics, not approximations.

4

Simulate Extreme Loading

LAMMPS molecular dynamics with the MACE universal potential simulates high-strain-rate compression, shock propagation, and thermal loading. Dynamic mechanical response and failure analysis — for any element combination, without pre-fit potentials.

5

Score Mission Fitness

Configurable scoring indices evaluate candidates across mechanical merit, weight, manufacturability, thermal stability, and cost. Candidates are ranked against known reference materials in application-specific benchmark sets tailored to each campaign.

6

Harden & Package Evidence

An 8-step hardening pipeline — tight relaxation, DFT elastic tensor, multistructure polymorph search, MD simulation, phonon stability, thermal AIMD, evidence bundle, benchmark ranking — produces cryptographically signed evidence packages ready for contract delivery and patent filing.

Live System Preview

This is what autonomous discovery looks like.

Real-time dashboard running on dedicated GPU infrastructure. Compositions shown below are simulated examples — actual discoveries are proprietary.

Live
Overview Performance DFT & Discovery Intelligence Hardening Results
Active Learning Cycle Each iteration follows this pipeline
Generate
GA-CSP + CompExplorer
ML Screen
MACE + ALIGNN + MatterSim
8-Obj Pareto
Hardness, ACI, Stability
DFT Validate
Quantum ESPRESSO GPU
6 AI Agents
Analyze & Steer
Retrain
Learn from DFT
Running Live Activity
Iteration 7 / 3 — 46%
Generating & screening candidates... 52s MP validating 258 candidates... 0s Running Pareto optimization + UQ...

45

Obs This Iter

250

Total Obs

22

DFT Validated

0.23

Rank ρ (Spearman)

Plausibly Novel Discoveries 12

Compositions that cleared all internal checks and are not in any known database or structural family. Strongest IP candidates — pending formal prior-art search. Only compounds with formation energy < 0 eV/atom appear here.

# Composition Hardness (GPa) E_form Ehull Crystal ACI Source Status
1 ZrHfBN2 31.4 -3.217 eV/atom 8 meV P6₃/mmc 72 Iter 5 HIGH_PRIORITY_IP
2 TiAlCrN3 28.7 -2.943 eV/atom 14 meV Fm3̅m 67 Iter 4 HIGH_PRIORITY_IP
3 SiB4C2N 34.1 -2.811 eV/atom 21 meV R3̅m 58 Iter 6 HIGH_PRIORITY_IP
4 WMoTiC3 26.3 -2.654 eV/atom 43 meV Pm3̅m 54 Iter 3 HIGH_PRIORITY_IP
5 HfTaNbN2 29.8 -2.519 eV/atom 11 meV P4/mmm 63 Iter 5 HIGH_PRIORITY_IP
6 VCrAlBN3 25.1 -2.387 eV/atom 37 meV I4/mmm 51 Iter 7 HIGH_PRIORITY_IP
7 NbZrSiC2O 22.9 -2.201 eV/atom 52 meV Cmcm 47 Iter 2 HIGH_PRIORITY_IP
Agent Guidance Active From Iter 6

10

Element Boosts

12

Plausibly Novel

2

Acquisition Targets

3

Confirmed Hypotheses

Ti +80% Al +80% Zr +80% N +80% O +80% W +80% Mo +80% Cu +80% H +30% As +30%

Simulated data shown for demonstration. Actual discoveries are proprietary and under IP review.

Why UnderationAI

Generate. Screen. Validate. Rank. Repeat.

Six generators propose novel crystal structures. An ML ensemble (MACE, ALIGNN, MatterSim, MEGNet) screens in seconds. GPU-accelerated Quantum ESPRESSO validates with DFT physics. LAMMPS simulates extreme mechanical and thermal loading. A 6-objective Pareto optimizer decides what to explore next. Every discovery gets a cryptographically signed evidence bundle. The entire loop runs autonomously on dedicated GPU hardware — no human in the loop, no cloud dependencies.

6 Crystal Structure Generators

GA-CSP evolves structures across 8 crystal prototypes (HCP, BCC, FCC, corundum, spinel, rutile, rocksalt, B4C-rhomb). A composition explorer samples from a configurable palette of defense-relevant elements. A substitution engine mutates DFT-confirmed hits. HEA, DiffCSP, and MatterGen generators provide additional diversity. Every iteration proposes 32 novel candidates.

GPU-Accelerated DFT Validation

Quantum ESPRESSO runs on an RTX 5090 with CUDA Fortran — full PBE relaxation, spin-polarized calculations for magnetic transition metals, formation energy, elastic tensor, and bulk/shear/Young's modulus. Each DFT calculation provides ground-truth physics that trains the ML models to be more accurate next iteration.

Auditable Evidence Bundles

Every hardened candidate gets a cryptographically signed evidence package: SHA-256 content hashes, Ed25519 digital signatures, per-metric provenance labels (DFT-computed, ML-predicted, elastic-derived), readiness scores, and evidence confidence ratings. Contract-deliverable and patent-ready.

Molecular Dynamics & Mechanical Scoring

LAMMPS with the MACE-MP-0 universal potential simulates extreme mechanical and thermal loading for any element combination — no pre-fit potentials needed. Dynamic compressive strength, failure analysis, and configurable scoring indices answer “will this survive the mission environment?” computationally, before any lab time.

5-Tier Novelty Gate

Before spending GPU time on DFT, every composition is checked against 12M+ entries across Materials Project, AFLOW, OQMD, NOMAD, arXiv, and CrossRef. Structural family classification catches obvious variants. Only compositions that survive all 5 tiers proceed to validation — we only pursue materials that are genuinely novel.

On-Premise & ITAR-Ready

The entire stack — ML models, DFT engine, LLM agent, knowledge store — runs on dedicated GPU hardware with no cloud dependencies. The local LLM provides agent guidance (element boosts, hypothesis testing) without sending data to third-party APIs.

Results to Date

This isn't a prototype. It's running.

The platform has been running autonomously on dedicated GPU hardware since early 2026. These are real numbers from a live system — not projections.

9

Novelty-Screened
Candidate Compounds

No matches in Materials Project, OQMD, AFLOW, or open literature.

130+

Autonomous Discovery
Iterations Completed

Each iteration: generate, screen, validate, learn, repeat.

GPU

Accelerated DFT
on RTX 5090

CUDA Fortran QE. 4 parallel DFT jobs. 99% GPU utilization.

6

Structure Generators
Running in Parallel

GA-CSP, HEA, composition explorer, substitution, DiffCSP, MatterGen.

Live Pipeline Performance

191+

DFT observations
recorded

8

Crystal prototypes
explored

28

Defense-relevant
elements

12M+

Known compounds
checked per candidate

Evidence System

Every prediction is auditable.

We don't just discover materials — we produce defensible evidence packages suitable for contract deliverables, patent filings, and defense prime technical review.

What's in an Evidence Bundle

  • Crystal structure (CIF format)
  • Formation energy & energy above hull (DFT-computed)
  • Elastic tensor: bulk, shear, Young's modulus (DFT stress-strain)
  • Elastic-derived hardness estimate (Chen/Tian model)
  • Dynamic compressive strength (LAMMPS impact MD)
  • Phonon stability analysis
  • Thermal AIMD stability (Lindemann criterion)
  • ACI score with component breakdown
  • Benchmark ranking vs known reference materials
  • SHA-256 hash + Ed25519 digital signature

Trust Tiers

T0 ML Screening

Rapid neural network predictions. Broad coverage, explicitly provisional. Use for initial down-selection.

T1 DFT Validated

First-principles physics confirmation. Formation energy, stress tensor, and equilibrium structure verified by quantum mechanics.

T2 Hardening Complete

Full 8-step pipeline: DFT elastic tensor, multistructure polymorph search, impact MD, phonon, thermal AIMD, evidence bundle, benchmark ranking. Cryptographically signed. Patent-ready.

Every metric is labeled with its origin: DFT-computed, ML-predicted, elastic-derived, or MD-derived. You always know exactly how each number was produced.

How We Work With You

You define the requirement. We deliver candidates.

Our discovery platform is material-class agnostic. Give us a target property envelope and we run autonomous computational campaigns that produce ranked, validated candidate materials with full documentation.

Step 1

Define the Envelope

You specify target properties — hardness, density, thermal limits, radar characteristics, cost constraints. We translate them into computational screening thresholds.

Step 2

Autonomous Campaign

Our platform generates thousands of candidates, screens with ML ensembles, validates top hits with first-principles DFT, and ranks by feasibility — all in a closed loop.

Step 3

Validated Shortlist

You receive 3–5 high-confidence candidate materials with predicted properties, uncertainty estimates, DFT validation data, and synthesis feasibility assessments.

Step 4

Transition to Lab

Full technical reports, candidate datasets in standard formats (CIF, JSON), and synthesis route recommendations ready for experimental validation by your team or ours.

SBA-Certified SDVOSB — Active on DARPA & AFRL BAAs

AI-driven materials discovery as a research service.

Give us a target property envelope. We deliver ranked candidate materials with DFT-validated properties, calibrated uncertainty estimates, and synthesis feasibility assessments — in weeks, not years. Your requirements in, novel material candidates out.

Active & Target Engagements

Pursuing defense materials research contracts.

DARPA DSO

Autonomous multi-fidelity materials discovery for rapid response to emerging defense material requirements

Executive Summary Submitted

AFRL Materials & Manufacturing

AI-accelerated ceramic materials discovery for extreme-environment survivability and protective applications

POC Engagement In Progress

Navy SBIR — NV028

Overlay and bond coatings for gas turbine hot corrosion resistance using ICME methodology

Proposal In Development

About the Founder

Built by someone who doesn't accept “good enough.”

I grew up in Rittman, Ohio — a small town surrounded by industries that couldn't adapt and eventually faded away. I refuse to let that happen to our defense industrial base.

As a 9½-year USSF veteran with 12 years in the space community, I've operated spacecraft in the most demanding 24/7 environments on the planet. As a Mission Operations Lead, I've logged 1,400+ console flight hours, executed 6,000 satellite contacts, and resolved hundreds of anomalies in real time.

I built this platform the same way I learned to operate weapon systems — hands-on, asking hard questions until the system delivers exactly what the mission demands. Every line of code, every DFT pipeline, every neural network integration — built from scratch by one person who believes our defense forces deserve better materials.

Based in Hazel Green, Alabama — running real GPU inference on dedicated hardware, discovering materials that don't exist in any database in the world. Underation Federal Services LLC is an SBA-certified Service-Disabled Veteran-Owned Small Business (SDVOSB) actively engaged with DARPA, AFRL, and the Navy SBIR program.

USSF Veteran SBA-Certified SDVOSB 1,400+ Flight Hours 12 Years Space Community

Get in Touch

Let's talk about what you need.

Whether you're a defense program office, a prime contractor, a materials manufacturer, or a research lab — we'd like to hear from you.

Opens your email client. We respond within 24 hours. All inquiries are confidential.

Open-Source Attribution

UnderationAI is built on the shoulders of world-class open-source projects. We gratefully acknowledge the following software, models, and data sources that power our platform.

Machine Learning Models & Pretrained Weights

MACE-MP-0 MIT

Universal interatomic potential for energy, force, and stress prediction. Batatia et al., University of Cambridge.

github.com/ACEsuit/mace
ALIGNN NIST Public Domain

Atomistic line graph neural network for materials property prediction. Choudhary et al., NIST.

github.com/usnistgov/alignn
MatterSim MIT

Formation energy prediction and structure relaxation. Microsoft Research.

github.com/microsoft/mattersim
M3GNet / MEGNet (MatGL) BSD-3-Clause

Graph neural network interatomic potentials for hardness and modulus prediction. Materials Virtual Lab.

github.com/materialsvirtuallab/matgl
DeepSeek-R1 DeepSeek License

Large language model powering AI advisory agents for discovery, validation, and IP scouting. DeepSeek AI.

huggingface.co/deepseek-ai
CDVAE MIT

Crystal Diffusion Variational Autoencoder for generative structure design. Xie et al.

github.com/txie-93/cdvae

Simulation & Physics Software

Quantum ESPRESSO GPL v2

Density functional theory for formation energy, elastic tensors, phonon stability, and AIMD. Giannozzi et al.

quantum-espresso.org
LAMMPS GPL v2

Large-scale molecular dynamics for impact simulation and shock propagation. Thompson et al., Sandia National Labs.

lammps.org
ASE LGPL-2.1

Atomic Simulation Environment for structure manipulation and calculator interfaces.

wiki.fysik.dtu.dk/ase
pymatgen MIT

Materials analysis, phase diagrams, and energy above hull computation. Ong et al.

pymatgen.org
phonopy BSD-3-Clause

Phonon calculations for dynamical stability analysis. Togo & Tanaka.

phonopy.github.io/phonopy
spglib BSD-3-Clause

Crystal symmetry finder and space group analysis.

spglib.readthedocs.io

ML Frameworks & Optimization

PyTorch BSD-3-Clause

Core deep learning framework.

pytorch.org
BoTorch MIT

Multi-objective Bayesian optimization (Pareto front).

botorch.org
e3nn MIT

Equivariant neural network library (used by MACE).

github.com/e3nn/e3nn
DGL Apache 2.0

Deep Graph Library for graph neural networks.

github.com/dmlc/dgl

Materials Data Sources

Materials Project CC BY 4.0

Reference data for known materials, phase diagrams, and energy above hull. Jain et al.

materialsproject.org
OQMD CC BY 4.0

Open Quantum Materials Database for novelty verification and formation energy reference.

oqmd.org
AFLOW CC BY 4.0

Automatic FLOW for bulk material data and novelty checks. Curtarolo et al., Duke University.

aflow.org
NOMAD CC BY 4.0

Novel Materials Discovery laboratory for novelty verification across 12M+ entries.

nomad-lab.eu
JARVIS (NIST) NIST Public Domain

Crystal structure tools and materials data. Choudhary et al., NIST.

jarvis.nist.gov
OpenAlex CC0

Open scholarly knowledge graph (~250M works) for literature novelty verification.

openalex.org

Infrastructure & Compute

NVIDIA CUDA NVIDIA EULA

GPU-accelerated computing for DFT, ML inference, and MD simulation.

developer.nvidia.com/cuda-toolkit
FastAPI MIT

High-performance web framework for REST API and dashboard backend.

fastapi.tiangolo.com
llama.cpp MIT

Local LLM inference engine with GPU offload. Gerganov et al.

github.com/ggerganov/llama.cpp
ChromaDB Apache 2.0

Vector database for agent semantic memory and retrieval-augmented generation.

github.com/chroma-core/chroma

Pseudopotentials & Reference Data

pslibrary (PBE Pseudopotentials) GPL

PBE-PAW/RRKJUS pseudopotentials for 49+ elements. Dal Corso, Comput. Mater. Sci. 95, 337 (2014).

dalcorso.github.io/pslibrary
PBE Exchange-Correlation Functional

Perdew, Burke & Ernzerhof, Phys. Rev. Lett. 77, 3865 (1996). Standard DFT functional used for all first-principles calculations.

All trademarks and registered trademarks are the property of their respective owners. The use of these open-source tools and data does not imply endorsement by their developers or institutions. UnderationAI complies with all applicable open-source licenses. Computational results are screening-level predictions and do not constitute certified material qualification. All predictions require experimental validation before real-world deployment.