From computation to discovery.

0.237 eV Band-Gap MAE on 1,048 Materials — Without Training on a Single One

A single fixed predictor on the same 1,048-material public cohort where modern graph-neural-network ML reports ~0.31–0.33 eV MAE — matching state-of-the-art accuracy at millisecond speed, with zero training data and zero fitted parameters. Now extends to the full electronic-property bundle: carrier concentration (99.0% within factor 3, 100% type accuracy), mobility & conductivity (100% within factor 3), and band-edge alignment (~88% Type I/II/III). Predictions hold for novel materials the predictor has never seen.

Head-to-Head with First-Principles DFT — 15 Canonical Materials

Local GPAW PBE screening setup (200 eV, 6³ k-points) on Si, Ge, GaAs, GaN, ZnO, MgO, TiO₂, NaCl, Al, Cu, Fe, Ni, graphite, h-BN, MoS₂. FluxMateria predicts lattice composition-only at 0.1% median off experiment, and beats this PBE setup on band gap (7.6% vs 45.1%) and magnetic moment (3.6% vs 9.0%) at ~25,000× the per-material wall time.

The FluxMateria Semiconductor Atlas

4.6 million semiconductor candidate combinations, enumerated and mapped by a physics-native engine in 199 seconds on a laptop. 87.5 % match against Madelung / NSM Ioffe / Sze references, with zero parameters fit to mobility or bandgap data. Existing platforms organize computed materials — FluxMateria generates candidate universes.

245 FDA Drugs. 8 ADMET Endpoints. One Call.

A unified mechanism-aware ADMET pipeline that replaces the four-to-six-tool stitched stack pharma teams currently maintain. Three strict #1 SOTA endpoints (solubility, metabolism, PPB-noise-floor) plus DILI AUROC 0.9597 on the comparable TDC binary task and Caco-2 MAE 0.277 matching the public TDC reference SOTA from pure physics. One input, one output schema, one mechanism-evidence trail per prediction. Annualized TCO materially below Schrödinger, Simulations Plus, and in-house ML pipelines.

5,008 DFT-Grade Property Predictions in 13.5 Seconds

313 materials × 16 properties, end-to-end. Accuracy audited against the materials-universal benchmark: 1.17% MAPE on family holdout — 9 to 31 times more accurate than AFLOW, JARVIS, and MatBench on the same hard split. Annualized total cost of ownership: materially below DFT HPC, in-house ML pipelines, and stitched commercial stacks. Real-time turnaround inside the design loop.

Alzheimer's Amyloid-Pathway Discovery

A strict BACE1-centered Alzheimer's workflow that forced CNS feasibility and explicit anti-target control against AChE and KCNH2. Final result: 48 strict passes from 707 candidates, two independent BACE1-native chemotypes, and a downloadable white paper.

Iptacopan (Fabhalta®) Factor B Profile — Predicted Blind

From SMILES alone and without training on this compound: 9 of 9 ADMET endpoints match the FDA label, CYP2C8 correctly identified as primary metabolic enzyme, Factor B / Factor D selectivity at 2.5 log units, and a physics-derived binding scorer ranks a 179-compound ChEMBL Factor B panel at Pearson 0.53 with zero training data and zero fitted parameters.

Better Battery Cathodes Through Interface-Aware Screening

A local battery workflow that showed why bulk-only ranking is not enough. Bulk chose LiNiO2, interface reopened LiMnPO4, battery-native scoring elevated LiMnO2, and prototype handoff selected Li4Ti5O12 as the best immediate build candidate.

Solar Absorber Discovery and CdTe Replacement Challenge

A two-phase solar workflow: first, stack-aware scoring moved the winner from InP to CdTe + ZnO + Mo; then a stricter replacement challenge surfaced CuS and CuSe as novel CdTe-replacement lanes.

Novel Superconductor Family: W-Cuprate Discovery

12,800 candidates screened to 3 experiment-ready leads. A novel tungsten-cuprate composition family with predicted T_c above 160 K, synthesis protocols, and a pending patent — delivered in under 24 hours.

Pure-Physics Caco-2 Permeability Matches Trained-ML State-of-the-Art

On the TDC caco2_wang scaffold-stratified test set, FluxMateria reaches MAE 0.277 log units versus the published trained-ML state-of-the-art of 0.276 — with zero Caco-2 training labels consumed at build time.

Would FluxMateria Have Caught These Clinical Failures?

Blind retrospective screening of 34 drugs withdrawn or failed due to ADMET toxicity. 88.2% detection rate with zero false positives on 16 safe controls — in 10.5 seconds total.

Predicting Drug MoA from Structure Alone — No ML Required

120 drug-target pairs across 12 target families. 95% accuracy on ChEMBL external validation. Agonist, antagonist, and inhibitor classification from SMILES — in 13ms per prediction.

56 Textbook Reactions Resolved in Under 28 Milliseconds

SN1, SN2, E1, E2, and E1CB mechanism prediction on textbook reactions from Clayden, Bruice, and March. 98.2% accuracy including condition-dependent flips — in sub-millisecond time per reaction.

Beating DFT on Curie Temperature — 107 Materials, 17 Families

4.59% MAPE across 107 magnetic materials from composition alone. DFT achieves 15–30% and takes hours. FluxMateria does it in milliseconds with zero fitted parameters.

16 Properties. All Under 1%. One Axiom. 2.7 Milliseconds.

Complete material characterization from a single formula. Band gap, lattice constants, hardness, thermal expansion, melting point, density, and 10 more — all under 1% error. 856 materials validated with 15 cross-validation scenarios.

Inverse Catalyst Discovery: Rediscovery and Exclusion Search

Specify activity, selectivity, stability, and cost constraints, then force the engine off the obvious incumbent lane when needed. FluxMateria converges to the same catalyst families industry and academia already trust, then moves into credible adjacent chemistry under exclusions. The hardened public benchmark scored 1,000 full-stack API calls in 19.7 seconds.