The physics-only approach to molecular prediction.
How FluxMateria derives molecular and material properties from first-principles geometry — without machine learning, without empirical fitting, without training data.
How FluxMateria derives molecular and material properties from first-principles geometry — without machine learning, without empirical fitting, without training data.
Most computational chemistry tools fall into two categories: quantum mechanical methods that solve the Schrödinger equation (accurate but slow), or machine learning models that learn statistical patterns from data (fast but fragile).
FluxMateria takes a third path. It derives molecular and material properties directly from molecular geometry and fundamental physical constants. No training datasets. No fitted coefficients. No neural networks.
The result is a physics engine that combines the traceability of first-principles methods with the speed needed for library-scale screening.
Molecular structure (SMILES, SDF) or material composition
Geometry analysis → Physical constants → Property derivation
Property predictions + confidence scores + full provenance chain
Three approaches to molecular property prediction. Different trade-offs.
Five properties that distinguish physics-only prediction from statistical modeling.
Predictions do not depend on historical datasets. There is no training set, no validation split, no risk of data leakage. The physics engine has never “seen” your molecules before — it derives their properties from structure.
Every parameter in the computation derives from fundamental physical constants and molecular geometry. There are no empirically tuned weights, no calibration factors, no regression coefficients.
Same input, same output, every time. Results are version-controlled and replayable. No stochastic sampling, no random seeds, no run-to-run variation.
Because predictions derive from physics rather than patterns, novel scaffolds and unexplored chemical space work on day one. There is no “applicability domain” in the ML sense — coverage is determined by physics, not training data.
Every prediction has a full provenance chain. You can trace any ADMET score, any band gap estimate, any barrier prediction back through the physics to understand exactly why that value was computed. This enables audit, review, and scientific challenge.
Every FluxMateria prediction carries a per-endpoint confidence score. This is not a statistical uncertainty estimate — it measures how well the underlying physics covers the structural features of the input.
The structural features are well-covered by the physics engine. Use these predictions for screening decisions.
Partial coverage. Use directionally, but prioritize experimental follow-up for critical decisions.
Physics coverage gap. Flag for experimental validation. Do not use as a pass/fail gate.
A single molecule can have high-confidence metabolism and low-confidence hERG. Confidence is assessed independently for each property endpoint.
Example: same molecule, different confidence per endpoint
Key benchmarks across four domains. Full methodology and raw data available on request.
| Domain | Key metric | Coverage | Speed |
|---|---|---|---|
| ADMET | 82.8% metabolism accuracy | 8 validated endpoints | ~350 mol/sec |
| Materials | 0.15 eV band gap MAE | 82 materials, 73 elements | 106× vs DFT |
| Mechanisms | 100% classification (SN1/SN2/E1/E2/E1cb) | 336 experimental cases | 106× vs DFT |
| Binding | 0.6 pKi MAE | 1,000+ targets, 500K+ molecules | ~50 ms/molecule |
No computational method covers everything. Here is where FluxMateria has boundaries — and how we handle them.
Some structural features are not yet fully covered by the physics engine: macrocycles, metallocenes, heavily fluorinated scaffolds, and molecules above ~800 Da. Confidence indicators flag these automatically.
FluxMateria is designed for screening and triage — reducing thousands of candidates to a validated shortlist. For final go/no-go decisions on your top 5 candidates, DFT or experimental validation remains appropriate.
Currently supports 73 elements (H through Bi). Lanthanides and transactinides are not yet supported. Materials coverage spans 14 crystal structure categories and 82 validated compounds.
Coverage expands with each release as new physics derivations are validated. If you encounter systematic gaps in your chemical series, we want to hear about it — it helps us identify the next physics to derive.
DFT methods solve the Schrödinger equation iteratively for each molecular configuration. This is computationally expensive — hours to days per molecule, depending on system size.
FluxMateria bypasses iterative quantum mechanical calculation entirely. Properties are derived analytically from molecular geometry and physical constants. The computation is deterministic and completes in milliseconds.
This is not an approximation of DFT. It is a fundamentally different computational approach that arrives at comparable accuracy through a different physical framework.
FluxMateria occupies the space between DFT accuracy and ML speed.
The best way to assess FluxMateria is to run it on your own molecules or materials. Start with a focused pilot or explore the published benchmarks.