Materials Lab
DFT-quality quantum chemistry and molecular dynamics via API. MACE-MP-0, Egret-1, AIMNet2, Orb-v3 — all neural network potentials, all under one API key.
What's inside
MACE-MP-0 Universal Potential
89-element coverage. Equivariant message passing architecture. Replaces DFT for molecular dynamics at 1000× the speed. MIT license.
Egret-1 DFT-Quality Properties
Neural network potential for organic molecules. Predicts pKa, logP, LogD7.4, BDE, and redox potentials at DFT accuracy in under 1 second.
AIMNet2 + Orb-v3
AIMNet2 for organic chemistry, Orb-v3 for 100K-atom systems. All three potentials (Egret / AIMNet / Orb) exposed via a single `method` parameter.
Fukui Indices & Reactivity
Per-atom electrophilic / nucleophilic / radical Fukui indices for reactivity analysis. Drop-in replacement for expensive CAM-B3LYP calculations.
FAQ
What is MACE-MP-0?›
MACE-MP-0 is a universal neural network interatomic potential trained on the Materials Project. It covers 89 elements and can simulate molecular dynamics for materials, catalysts, and inorganic systems at DFT-level accuracy but 1000× faster. Developed by the University of Cambridge ACEsuit team, MIT license.
What is Egret-1 and who uses it?›
Egret-1 is a neural network potential for small organic molecules, trained on a massive corpus of DFT calculations. It predicts pKa, logP, bond dissociation energies, and redox potentials at DFT quality in <1 second per molecule. Rowan Scientific popularized it as their flagship chemistry tool. SciRouter is the first aggregator to host it alongside bio models.
When should I use MACE vs Egret vs AIMNet?›
MACE-MP-0 for materials science (crystals, catalysts, inorganic). AIMNet2 for small organic molecules where you need pKa / logP. Egret-1 for drug-like molecules where you want DFT-quality binding energies. Orb-v3 for very large systems up to ~100K atoms.
Is this as accurate as real DFT?›
Modern NN potentials match DFT accuracy on the tasks they were trained for. For tasks far outside training distribution (exotic chemistry, unusual elements, high pressures), DFT is still more reliable. For routine drug-like molecules and common materials, NN potentials are faster, cheaper, and functionally equivalent.
What hobbyists actually run this?›
Weekend chemistry hackers use Egret-1 for instant pKa on TikTok-famous molecules. Battery researchers use MACE-MP-0 for rapid electrode screening. Materials postdocs use Orb-v3 for surface reactions. Rowan Scientific built their business around this — SciRouter makes the same models available under one API.
Are real models running or mock mode?›
Sprint 51 ships with deterministic mock mode. Real RunPod GPU workers activate as endpoint IDs are provisioned. Every response includes a 'dispatch_mode' field (mock | runpod) so you always know which backend served your call.