🤖 Machine Learning · Machine Learning Potentials
Methodology
Machine-learning potentials deliver near first-principles accuracy at a 10,000-fold speedup, making large-system, long-timescale dynamics possible. We handle the entire workflow, from dataset construction to potential validation.
Applications
Related projects
Potential training ★ Flagship
MACE / NequIP / DeepMD — flagship scenario
Material property prediction
Formation energy / band gap / elasticity
Inverse design
Composition from target properties
Case Studies
FAQ
We build our datasets with our own AIMD and DFT runs, iterated through active learning — we do not rely on public data of unknown provenance.
Before delivery we run triple validation on energies, forces and stresses, plus a benchmark report against a DFT test set; the error tolerance is written into the SLA.
Standard potentials are delivered in 14 days, including dataset construction and validation.
Includes the Computational Traceability Report: datasets, hyperparameters and validation reports are all archived.