Aether for materials science.
An AI scientist for the full materials lifecycle.
AetherMat-Sci inverts discovery: from target property to a physics-validated material. Sixteen specialist agents — inverse design, crystal-structure generation, ML-accelerated property prediction, simulation, microstructure vision, failure analysis, manufacturing readiness and qualification — under one model, with a calibrated uncertainty and an honest accuracy tier on every result.
- Agents
- 16
- Path
- Target → Qualified part
- Families
- Alloy · ceramic · polymer · battery
- Tiers
- Empirical → ML → MLIP → DFT
Discovery to qualification, under one model.
The incumbent materials stack is a dozen seat-licensed tools with a dozen file formats. AetherMat-Sci is one model and one project file — built so the inverse engine feeds the structure generator feeds the property model feeds the qualification report without a single CSV in between.
Materials knowledge graph
Composition → process → microstructure → property → failure → manufacturing, as a typed causal graph. Semantic + property + similarity search, JSON/Cypher export. The system's memory (SRS FR-13).
Inverse / generative design
Target property → composition. Multi-objective Pareto search across alloys, ceramics, polymers, battery and catalyst families, with regulatory + supply-chain + cost constraints and an honest yield funnel (SRS FR-8 / IME).
Crystal structure generator
Composition → atomic realization. Lattice/prototype, Wyckoff symmetry (spglib), defects, dopants, CSL grain boundaries, Miller slabs, CIF/POSCAR/XYZ — validated five ways (CSG module).
Property prediction engine
Trained ML ensemble on 30k Materials Project entries — formation energy and band gap with conformal-calibrated uncertainty — plus empirical surrogates and CHGNet MLIP. Tiered escalation (SRS FR-4/7).
Simulation copilot
Natural language → solver setup. Mesh/solver/BC recommendation, Latin-hypercube DOE, Morris sensitivity, optimization loops, and job decks for commercial CAE suites / commercial CAE suites / open-source CFD / MD packages (SRS FR-5).
Microstructure intelligence
Micrograph → engineering numbers. ASTM E112 grain size, porosity, crack and inclusion detection, phase fractions, pass/fail scoring, and reverse linkage to the process that caused it (SRS FR-6).
Processing → property
Causal physics: Hall-Petch, JMAK grain growth, additive-manufacturing energy density → porosity → fatigue. Inverse process design for a target property (SRS FR-7).
Failure analysis agent
Fractography + load + environment → ranked root cause with calibrated posteriors. Returns inconclusive when evidence is weak rather than forcing an answer (SRS FR-8).
Lab automation & active learning
Exact Gaussian-process Bayesian optimization with EI/UCB, batch proposal, stopping criteria, safety constraints, and machine-actionable experiment specs for self-driving labs (SRS FR-15).
Literature & patent mining
Text → structured composition / process / property facts with unit normalization, contradiction surfacing, dedup, and TF-IDF retrieval — never a fabricated citation (SRS FR-14).
Manufacturing readiness
Manufacturability scoring per route, process-window robustness, yield prediction, cost model, supply-chain and safety pre-checks, and a lab→pilot→production scale-up verdict (SRS FR-12 / IME).
Regulatory & qualification
Requirement mapping (ASTM/ISO/ASME/FDA), evidence aggregation, gap analysis, traceability lineage, and an audit-ready report that says “not certifiable yet” when evidence is missing (SRS FR-16).
Every number knows how good it is.
The governing principle: experimental, computed and model-predicted values are never conflated; uncertainty is always reported; a surrogate is never presented as DFT or experiment. The four-tier ladder makes that structural, not a promise.
Empirical surrogate
Closed-form physics (rule-of-mixtures, PREN, Labusch strengthening, Hall-Petch). Milliseconds. Screening only — always labelled as such.
Trained ML ensemble
Deep ensemble on Materials Project with conformal uncertainty. Formation-energy MAE 136 meV/atom, band-gap MAE 0.37 eV, ~90% interval coverage.
MLIP physics
Pretrained CHGNet — DFT-quality energy, forces and relaxation across the periodic table, on GPU, in seconds. The fast physics check before any DFT.
First-principles (escalation)
VASP / Quantum ESPRESSO / phonons / elastic tensors. Honest engine-hooks: a property we cannot yet compute returns no number with the engine it requires — never a fabricated value.
Five material families, family-specific rules.
Not one generic model — each family carries its own design physics.
Titanium, nickel superalloys, steels, aluminium. CALPHAD phase rules, γ′ fraction + PHACOMP TCP screen, Larson-Miller creep, PREN, density-by-rule-of-mixtures.
Perovskites and beyond — Goldschmidt tolerance factor + octahedral factor validity gates, sintering densification, thermal-barrier chemistry.
Fox/Flory-Fox Tg, group-contribution density and solubility parameter, crosslink and filler formulation for flexibility, heat resistance and biocompatibility.
Cathode/anode/electrolyte chemistry, theoretical gravimetric capacity, voltage and ionic-conductivity screening for LFP alternatives and solid-state.
Active-site + support + nanoparticle composition, surface-plane (100/110/111) generation for adsorption and selectivity.
Zinc-blende / diamond structure generation, dopant placement, band-gap prediction with the honest PBE-underestimation caveat.
Honest, measured numbers.
Trained and benchmarked on real Materials Project data. We report the numbers we actually measured — including where the model is weakest.
Formation energy · MP held-out (3k)
Band gap · MP held-out (3k)
Conformal intervals
Training set
V&V regression suite
RTM coverage
One model for the whole shelf.
A dozen seat-licensed tools and file formats collapse into one agent, one contract, one project file.
The whole lifecycle, on one model.
The materials stack is an informatics platform, a simulation suite and a database, each owning a slice. Aether covers discovery to qualification on one model — with calibrated uncertainty on every result.
R&D cycles, compressed.
What compounds is iterations per quarter, not any single solve being faster.
From knowledge graph to qualification — each emitting results under one scientific-integrity contract with stable contracts.
Empirical → trained ML → MLIP physics → first-principles. Every number carries its tier, provenance and calibrated uncertainty.
Materials Project, with conformal-calibrated uncertainty — and OQMD / AFLOW / MatBench in the ingestion pipeline.
An AI scientist, not a tool.
It plans, picks methods, runs them, validates against data, and writes the rationale — for an expert audience.
- 01
Reads the target
Ingests the property goals, the constraints (regulatory, supply-chain, cost) and the material family, and decomposes the brief into a typed study graph before any compute runs.
- 02
Picks the right tier
Knows when an empirical correlation is enough, when to call the trained ML ensemble, and when to escalate to CHGNet or DFT — by the uncertainty the conclusion needs and the budget available.
- 03
Generates and physics-filters
Inverse design proposes; Hume-Rothery / VEC / charge-neutrality and CHGNet relaxation dispose. Generative plausibility never overrides a failed physics check.
- 04
Validates against data
Cross-checks predictions against Materials Project and the literature. Conformal calibration keeps the stated 90% intervals honest; out-of-domain inputs are flagged, not silently extrapolated.
- 05
Writes the rationale
Each candidate ships with which targets it meets and by what margin, the trade-offs, the risks, and the recommended validation experiment — with the model-version hash.
- 06
Closes the loop
Proposed materials enter the active-learning queue as machine-actionable experiment specs; measured results return into the knowledge graph and sharpen the next round.
Materials intelligence that improves with every experiment.
Each measured result returns into the knowledge graph and the active-learning loop, so the model that designed a material gets better at designing the next one.
“It designed an alloy to our target instead of finding the closest one that already existed. That's a different kind of tool.”
Design a material against a target.
Open the studio to run property prediction, inverse design, crystal generation and knowledge-graph queries against the live agent — or request access to deploy it behind your firewall.
Prefer the raw API? The studio talks to a local agent at /materials/studio.